CN115049108A - Multitask model training method, multitask prediction method, related device and medium - Google Patents

Abstract

The embodiment of the specification discloses a multitask model training method, a multitask prediction method, a related device and a medium, wherein the training method comprises the following steps: the parameter weight of the sample feature in the attention module of the mth sub-model is determined, the parameter in the attention module of the (m + 1) th sub-model is determined according to the parameter weight in the attention module of the mth sub-model, and the multitask model can be trained based on the parameter weight in the attention module of the mth sub-model and the parameter in the attention module of the (m + 1) th sub-model. The multi-task model can enable each task in the multi-task to correspond to one sub-model, and the related information of the previous task is transferred to the next task through the attention module between the adjacent tasks, so that the next task can obtain a more accurate and more relevant prediction result by combining the related information of the previous task.

Description

Multi-task model training method, multi-task prediction method, related device and medium

technical field

The present specification belongs to the technical field of data processing, and in particular relates to a multi-task model training method, a multi-task prediction method, a related device and a medium.

Background technique

Generally, the way to predict a task can be obtained by inputting the data corresponding to the task into the trained model. This type of model can be obtained by training based on the learning mode of the task data, and when faced with complex tasks, it can be decomposed first and then Training is performed on multiple tasks respectively, and the final prediction result is obtained by combining the corresponding results of the multiple tasks.

Since there may be correlation between multiple tasks, the prediction result obtained by the traditional prediction method will lose the correlation information, so it is necessary to provide a technical solution with higher accuracy of the prediction result.

SUMMARY OF THE INVENTION

The embodiments of this specification provide a multi-task model training method, a multi-task prediction method, a related device and a medium, the technical solutions of which are as follows:

In the first aspect, the embodiments of this specification provide a multi-task model training method. The multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module. The method includes:

Determine the parameter weight of the sample feature in the attention module of the mth sub-model; where m is a positive integer less than M;

The parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model; wherein, the sample features include sample user features, sample product features, and the user's correspondence to the m+1 th sub-model. sample results of the task;

The multi-task model is trained based on the sample features, the parameter weights in the attention module of the mth submodel, and the parameters in the attention module of the m+1th submodel.

In the second aspect, the embodiments of this specification provide a multi-task prediction method, the method is applied to a multi-task model, the multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module respectively, Methods include:

Determine the parameter weight of the target feature in the attention module of the mth sub-model; where m is a positive integer less than M;

The parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model; wherein, the target features include target user features and target product features;

Based on the target feature and the parameters in the attention module of the m+1 th sub-model, the prediction result of the task corresponding to the m+1 th sub-model is obtained.

In a third aspect, the embodiments of this specification provide a multi-task model training device. The multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module. The multi-task model training device include:

The first processing module is used to determine the parameter weight of the sample feature in the attention module of the mth sub-model; wherein, m is a positive integer smaller than M;

The second processing module is configured to determine the parameters in the attention module of the m+1 th sub-model according to the parameter weights in the attention module of the m th sub-model; wherein the sample features include sample user features, sample product features, and user's The sample result of the task corresponding to the m+1th sub-model;

The training module is used for training the multi-task model based on the sample features, the parameter weights in the attention module of the mth sub-model, and the parameters in the attention module of the m+1th sub-model.

In a fourth aspect, the embodiments of this specification also provide a multi-task model training device, which may include: a processor and a memory; wherein, the memory stores a computer program, and the computer program is adapted to be loaded by the processor and execute the above-mentioned multi-task model Training method steps.

In a fifth aspect, the embodiments of this specification provide a computer storage medium, where the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above-mentioned steps of the multi-task model training method.

In a sixth aspect, an embodiment of this specification provides a multi-task prediction device, the device is applied to a multi-task model, the multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module respectively, The device includes:

The third processing module is used to determine the parameter weight of the target feature in the attention module of the mth sub-model; wherein, m is a positive integer smaller than M;

The fourth processing module is used to determine the parameters in the attention module of the m+1 th sub-model according to the parameter weight in the attention module of the m th sub-model; wherein, the target features include target user features and target product features;

The prediction module is used to obtain the prediction result of the task corresponding to the m+1 th sub-model based on the target feature and the parameters in the attention module of the m+1 th sub-model.

In a seventh aspect, embodiments of the present specification further provide a multi-task prediction apparatus, which may include: a processor and a memory; wherein, the memory stores a computer program, and the computer program is adapted to be loaded by the processor and execute the above-mentioned multi-task prediction method step.

In an eighth aspect, the embodiments of the present specification further provide a computer storage medium, where the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above-mentioned steps of the multi-task prediction method.

The beneficial effects brought by the technical solutions provided by some embodiments of this specification include at least:

In one or more embodiments of this specification, when training the multi-task model, the parameter weights of the sample features in the attention module of the m-th sub-model may be determined first, and according to the attention module of the m-th sub-model The parameter weight of the parameter determines the parameters in the attention module of the m+1th sub-model, and can be based on the parameter weight in the attention module of the m-th sub-model and the parameters in the attention module of the m+1-th sub-model to many The task model is trained. The multi-task model can correspond each task in the multi-task to a sub-model, and transfer the associated information of the previous task to the next task through the attention module between adjacent tasks, so that the next task can combine the The correlation information of the previous task can get more accurate and more relevant prediction results.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the drawings used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

1 is a schematic diagram of a prediction flow of an existing multi-task model provided by an embodiment of the present specification;

2 is a schematic diagram of a prediction flow of a multi-task model provided by an embodiment of the present specification;

3 is a schematic flowchart of a multi-task model training method provided by an embodiment of the present specification;

4 is a schematic structural diagram of a sub-model provided by an embodiment of the present specification;

FIG. 5 is a schematic structural diagram of a multi-task model provided by an embodiment of the present specification;

6 is a schematic flowchart of a multi-task prediction method provided by an embodiment of the present specification;

FIG. 7 is a schematic structural diagram of a multi-task model training device provided by an embodiment of the present specification;

FIG. 8 is a schematic structural diagram of a multi-task prediction apparatus according to an embodiment of the present specification;

FIG. 9 is a schematic structural diagram of another multi-task model training device provided by an embodiment of the present specification;

FIG. 10 is a schematic structural diagram of another multi-task prediction apparatus provided by an embodiment of the present specification.

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present specification.

The terms "first", "second", "third" and the like in the description and claims of the present application and the above drawings are used to distinguish different objects, rather than to describe a specific order. Furthermore, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally also includes For other steps or units inherent to these processes, methods, products or devices.

When a general life application is used by the user, it will provide the user with a recommendation service that can improve the user's needs according to the user's actions. Taking a third-party application that can perform payment functions as an example here, the user can browse and click on the relevant information of interest on the recommendation interface of the third-party application, and can make payment for the relevant information of interest. The whole process can be divided into multiple tasks, for example, when the user browses the recommendation interface of the third-party application, it can correspond to the exposure task of the relevant information that the user is interested in. The related information of interest may correspond to the click task of the related information of the user's interest, and the user may correspond to the purchase task of the related information of the user's interest when the user pays to purchase the related information of interest in the third-party application. It is understandable that the above-mentioned exposure tasks, click tasks and purchase tasks have sequential dependencies, that is to say, the click task can continue to occur after the exposure task occurs, and the purchase can continue to occur after the click task occurs. tasks, and there is a relationship between adjacent tasks.

Of course, the number of the multiple tasks corresponding to the actions performed by the user may not be limited to the three mentioned above. For example, for a purchase-type third-party application, when the user places an order in the purchase-type third-party application, the user can first browse in the purchase-type third-party application, and then click on the item that meets the needs, and then click on the item that meets the needs. Receive a deduction coupon from the corresponding store for the item that meets the demand, and then write off the item that meets the demand, and place an order for the item that meets the demand after the write-off. Among them, when the user browses the third-party application of the purchase type, it can correspond to the exposure task of the item that meets the user's needs, and when the user clicks on the item that meets the user's needs in the third-party application of the purchase type, it can correspond to the item that meets the user's needs. For the click task of an item, when a user receives a coupon in the store corresponding to the item that meets the demand, it can correspond to the task of receiving the coupon for the item that meets the user's demand. The write-off can correspond to the write-off task of the item that meets the user's needs, and when the user places an order after the write-off of the item that meets the user's needs, it can correspond to the task of placing the item that meets the user's needs.

Based on this, in order to obtain the prediction result of the user's demand, a model learning method can generally be used to learn and train multiple tasks corresponding to the user's demand, and the final prediction result can be obtained by combining the prediction results of each task.

Specifically, as a common way of predicting multiple tasks, the multiple tasks can be divided into multiple tasks first, the multiple tasks are learned separately, and the prediction results corresponding to the multiple tasks can be combined by combining the respective prediction results of the multiple tasks. to get the final prediction result of multi-task. Among them, each task can be learned and trained through the task model. For example, the sample feature vector corresponding to each task can be input into the task model for training, so that the feature vector corresponding to each task can be input into the trained task model. , and get the corresponding prediction results. It is understandable that a task model can be set for each task this time to ensure the accuracy of the prediction results of each task model. Here, the multi-task can be divided into three tasks, A, B, and C as an example. Input task A into the trained task model to obtain the corresponding prediction result a, and input task B into the trained task model. The corresponding prediction result b can be obtained, and the corresponding prediction result c can be obtained by inputting the task C into the trained task model. The final prediction result of the multi-task can be obtained by combining the prediction result a, the prediction result and the prediction result c. , such as but not limited to a×b×c.

Here, reference may also be made to a schematic diagram of a prediction flow of an existing multi-task model provided by an embodiment of the present specification shown in FIG. 1 . As shown in Figure 1, taking the multi-task model including two sub-models, and the multi-task corresponding to the event including the first task and the second task as an example, the prediction result of the first task can be expressed as the predicted probability of the event being clicked, the second task The prediction result of the event can be expressed as the predicted probability of being converted if the event is clicked, and the prediction result of multi-tasking can be expressed as the predicted probability of the event being clicked and then converted. Specifically, when predicting the occurrence probability of a multi-task corresponding to an event, the feature vector corresponding to the multi-task may be input into the first model to obtain a first prediction result corresponding to the first task. The first model may be a sub-model of the multi-task model, which corresponds to the task model of the first task, and can be obtained by training with sample features of known prediction results. Then, the feature vector corresponding to the multitasking can be input into the second model to obtain a second prediction result corresponding to the second task. Wherein, the second model may be a sub-model of the multi-task model, which corresponds to the task model of the second task, and can be obtained by training with sample features of known prediction results. Further, the target prediction result of the multi-task can be obtained by combining the first prediction result corresponding to the first task and the second prediction result corresponding to the second task, for example, by multiplying the first prediction result corresponding to the first task by A second prediction result corresponding to the second task is obtained. It can be understood that the first prediction result corresponding to the first task and the second prediction result corresponding to the second task are not limited to the sequence, for example, the second prediction result corresponding to the second task can also be obtained first. Then, the first prediction result corresponding to the first task is obtained, or the first prediction result corresponding to the first task and the second prediction result corresponding to the second task are simultaneously obtained.

It can be seen that the above-mentioned method of multi-task prediction does not take into account the strong correlation between each task, and only obtains the final prediction result of the loss of relevant information through the simple method of scalar multiplication, which is easy to affect the final prediction result. The accuracy of the forecast results.

Next, in order to better solve the above-mentioned technical problems, one or more embodiments of this specification will be explained.

Referring to FIG. 2, FIG. 2 shows a schematic diagram of a prediction flow of a multi-task model provided by an embodiment of the present specification. As shown in FIG. 2 , the multi-task model may include two sub-models (which may be represented as a first model and a second model), and the corresponding events may include the first task and the second task, where the first model may correspond to To predict the probability of occurrence of the first task, you can refer to the above-mentioned prediction result of the first task, which can be expressed as the predicted probability of the event being clicked, where the second model can correspond to the probability of predicting the occurrence of the second task. The obtained prediction result of the second task can be expressed as the predicted probability that the event is converted if it is clicked, and the predicted result of the multi-task can be expressed as the predicted probability that the event is clicked and then converted.

When predicting the occurrence probability of the multitasking corresponding to the event, the feature vector corresponding to the multitasking can be input into the first model to obtain the parameter weight in the attention module of the first model. The first model may include a first module and an attention module. Specifically, the feature vector corresponding to the multi-task may be input into the first module to obtain the first transformation information, and then the first transformation information may be input into the first module. To the attention module, the parameter weights in the attention module of the first model are obtained. It can be understood that the parameter weights in the attention module of the first model here can be used to represent the parameters corresponding to the association information between the first task and the second task, which can also correspond to the attention of the first model. A partial parameter of all parameters in the module. Here, all parameters in the attention module of the first model can be represented as A, B, C, D and E as an example, the parameter weights in the attention module of the first model can be represented as (0, 1 but not limited to) , 1, 0, 0), that is, the parameters corresponding to the association information between the first task and the second task include B and C.

The feature vectors corresponding to multitasking mentioned here may include, but are not limited to, user features and task product features, wherein the user features can be understood as the feature information of the target user to perform the event, such as the target user's identity information, specifically. It can include any at least one kind of information such as user name, user classification or user address. Here, taking the execution of an event applied to a purchase-type third-party application as an example, the characteristic information of the target user can be determined by querying the user information filled in by the target user. The characteristic information of the target user is in addition to the user name, user classification mentioned above In addition to the user address, the target user's browsing records or purchase records within a preset time may also be included, and exchange coupons or commodity coupons that the user has received may also be recorded, which is not limited in this embodiment.

Among them, the task product feature can be understood as the feature information used to represent the product corresponding to the event, such as the product name of the event, product production information, or any at least one kind of information such as product definition information. The product definition information can be understood as being used for Information that characterizes the functionality of the product. Here, the executed event is applied to a purchase-type third-party application as an example. When a certain product to be purchased is determined, the characteristic information corresponding to the certain product can be inquired at the store where the certain product is located, such as the products mentioned above. Information such as name, product production time, product function, etc., and preferential information corresponding to a certain product may also be recorded, which is not limited in this embodiment.

Further, after obtaining the parameter weights in the attention module of the first model, the parameter weights in the attention module of the first model can be transferred to the attention module of the second model, so that according to the parameter weights in the attention module of the first model The parameter weights in the attention module of one model determine the parameters in the attention module of the second model. The second model may include a second module and an attention module, and the attention module of the second model may adjust its own parameters according to the received parameter weights in the attention module of the first model, so that the adjusted parameters can be adjusted. The parameters in the attention module of the second model have the association information between the first task and the second task. Here, the parameters in the attention module of the second model can be expressed as B, D, E and F, and the parameters corresponding to the parameter weights in the attention module of the first model can be expressed as B and C as an example. The parameters in the attention module of the two-model can be denoted as B, C, D, E, and F. It can be understood that the structure of the attention module of the second model here may be consistent with the structure of the attention module of the first model, but the corresponding parameters are different.

It can also be understood that, in the process of transferring the parameter weights in the attention module of the first model to the attention module of the second model, the transfer can be carried out through the connection module but is not limited to, for example, through the fully connected layer. The transfer process is implemented to ensure the integrity of the parameter weights in the attention module of the first model during the transfer process.

Further, after the parameters in the attention module of the second model are determined, the feature vector corresponding to the multi-task can be input into the second model, and the attention module of the second model can predict the event corresponding to the event. Probability of multitasking. The second model mentioned above includes a second module and an attention module. Specifically, the feature vector corresponding to the multi-task can be input into the second module to obtain the second transformation information, and then the first The second transformation information is input into the attention module whose parameters have been adjusted, and the occurrence probability of the multi-task corresponding to the event is directly obtained. It can be understood that the prediction result obtained according to the second model is combined with the association information between the first task corresponding to the first model and the second task corresponding to the second model. The prediction results of the corresponding tasks of the sub-model, and then the final multi-task prediction results are obtained by combining the prediction results of each corresponding task. The entire prediction process only needs to obtain the final prediction result, which not only reduces the data processing time, but also obtains the final results. The correlation information between adjacent tasks is retained in the prediction result, so that the final prediction result is more accurate.

Of course, this embodiment can be, but is not limited to, a multi-task model including two sub-models. Here, the above-mentioned multi-task may include exposure task, click task, coupon task, write-off task and next task as an example. The model can also be set to include five sub-models (which can be represented as a first model, a second model, a third model, a fourth model, and a fifth model, respectively), wherein the first model can correspond to the probability of predicting the occurrence of the exposure task, The second model can correspond to the occurrence probability of the predicted click task, the third model can correspond to the occurrence probability of the coupon collection task, the fourth model can correspond to the occurrence probability of the write-off task, and the fifth model can correspond to the occurrence probability of the next task , and each model can include an attention module. Specifically, the feature vector corresponding to the multi-task can be input into the first model first to obtain the parameter weight in the attention module of the first model, and the parameter weight in the attention module of the first model can be transferred to the attention module of the second model to adjust the parameters in the attention module of the second model. Further, the feature vector corresponding to the multi-task can be input into the second model again to obtain the parameter weight in the attention module of the second model, and the parameter weight in the attention module of the second model can be passed. to the attention module of the third model to adjust the parameters in the attention module of the third model. Further, the feature vector corresponding to the multi-task can be input into the third model again to obtain the parameter weight in the attention module of the third model, and the parameter weight in the attention module of the third model can be passed. to the attention module of the fourth model to adjust the parameters in the attention module of the fourth model. Further, the feature vector corresponding to the multi-task can be input into the fourth model, the parameter weight in the attention module of the fourth model can be obtained, and the parameter weight in the attention module of the fourth model can be passed to The attention module of the fifth model to adjust the parameters in the attention module of the fifth model. Further, the feature vector corresponding to the multi-task can be input into the fifth model, and the final prediction result can be obtained through the attention module of the fifth model.

Please refer to FIG. 3, which shows a schematic flowchart of a multi-task model training method provided by an embodiment of the present specification.

As shown in Figure 3, the multi-task model training method may include at least the following steps:

Step 302: Determine the parameter weight of the sample feature in the attention module of the mth sub-model.

When the occurrence probability of multitasking needs to be predicted, the target feature corresponding to the multitasking can be input into the trained multitasking model mentioned in this embodiment, so as to obtain a more accurate prediction result. Based on this, the process of training the multi-task model mentioned in this embodiment is particularly important.

Specifically, before training the multi-task model, the number of tasks in the multi-task corresponding to the sample features and the sample results corresponding to each task may be determined. The sample features may include sample user features, sample product features, and user sample results for each task, and sample user features can be understood as sample feature information of the user who wants to execute the event, such as the user's sample identity information, which may specifically include the user Any at least one kind of information such as name, user classification or user address. The sample product feature can be understood as the sample feature information used to characterize the product corresponding to the event, such as the sample product name of the event, sample product production information, or sample product definition information, etc. Any at least one kind of information, the sample product definition information can be understood Information used to characterize product functionality. It can be understood that the requirements for sample products can correspond to multiple tasks, and each task can correspond to a known sample result. The sample result can be, but not limited to, when it is determined that the corresponding task does not occur. It is represented as character 0, and when it is determined that the corresponding task has occurred, it can be represented as character 1.

After determining the number of tasks in the multi-task corresponding to the sample features and the sample results corresponding to each task, the attention module of the model corresponding to the m-th task in the multi-task can be obtained according to the sample user features and the sample product features. parameter weights. The multi-task model in this embodiment may include M sub-models, each sub-model may include an attention module, each task in the multi-task corresponding to the sample feature may correspond to a sub-model, and the adjacent tasks can correspond to two adjacent sub-models. For example, the first task in the multi-task corresponding to the sample feature may correspond to the first sub-model in the multi-task model, and the second task in the multi-task corresponding to the sample feature may correspond to the second sub-model in the multi-task model sub-model, the m-th task in the multi-task corresponding to the sample feature can correspond to the m-th sub-model in the multi-task model, that is, the attention of the model corresponding to the m-th task in the multi-task mentioned above The parameter weight in the module can be understood as the parameter weight in the attention module of the mth sub-model in the multi-task model. Here, M may be a positive integer greater than m, for example, when m is 2, M may be a positive integer greater than 2.

It can be understood that the m-th task in the multi-task corresponding to the above-mentioned sample features can correspond to the m-th sub-model in the multi-task model, and the parameter weights in the attention module in the m-th sub-model can be used to represent The parameter corresponding to the association information between the m th task and the m+1 th task may also correspond to some parameters of all parameters in the attention module of the sub-model corresponding to the m th task. Here, all parameters in the attention module of the sub-model corresponding to the m-th task can be represented as A, B, C, D and E as an example, the parameter weights in the attention module of the sub-model corresponding to the m-th task It can be expressed as (0, 1, 1, 0, 0) but is not limited to, that is, the parameters corresponding to the association information between the m th task and the m+1 th task include B and C.

It can also be understood that the task corresponding to the mth sub-model in this embodiment may be any task from the first task to the penultimate task in the multi-task corresponding to the sample feature. Possibly, when the number of multi-tasks corresponding to the sample feature is 5, the value of m may be 1 or 2 or 3 or 4. Possibly, when the number of multi-tasks corresponding to the sample feature is 2, the value of m can only be 1.

Step 304: Determine the parameters in the attention module of the m+1 th sub-model according to the parameter weights in the attention module of the m th sub-model.

Specifically, after the parameter weight in the attention module of the mth sub-model is determined, the parameter weight in the attention module of the m-th sub-model can be transferred to the attention module of the adjacent m+1-th sub-model , to adjust the parameters in the attention module of the m+1th sub-model. Here, the parameters in the attention module of the m+1th model can be expressed as B, D, E and F, and the parameters corresponding to the parameter weights in the attention module of the mth model can be expressed as B and C as an example. After adjustment The parameters in the attention module of the m+1 th model can be denoted as B, C, D, E, and F.

It can be understood that the attention modules included in each sub-model in the multi-task model of this embodiment may adopt the same structure, and the difference lies in that the parameters of the attention modules included in each sub-model are different.

Step 306 , train the multi-task model based on the sample features, the parameter weights in the attention module of the m th sub-model, and the parameters in the attention module of the m+1 th sub-model.

Specifically, after the parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model, the sample user features and the sample product features in the sample features can be input into the m th sub-model. In the +1 sub-model, the occurrence probability of the m+1-th task is obtained through the attention module of the m+1-th sub-model. It can be understood that the occurrence probability of the m+1 th task obtained in this way includes the correlation information between the first m tasks, which is more accurate than the occurrence probability of the m+1 th task obtained in the prior art. higher.

Further, after obtaining the occurrence probability of the m+1 th task, the m th sub-model and the m+1 th sub-model of the multi-task model can be parameterized in combination with the sample results of the m+1 th task in the sample features. The training is optimized so that the occurrence probability of the m+1 th task is closer to the sample result of the m+1 th task in the sample features. It can be understood that in the training process of the multi-task model, as the parameters in the m-th sub-model change, the parameter weights in the attention module of the m-th sub-model also change, and then the m+1-th sub-model changes. The parameters in the attention module of the model will also change, which in turn can cause the occurrence probability of the m+1 th task to be closer to the sample result of the m+1 th task in the sample features.

It should be noted that when m is any positive integer greater than 1, the parameter weights in the attention modules of the corresponding sub-models can be obtained in turn according to the arrangement order of each task, and then the attention of the m-th sub-model can be obtained. The parameter weights in the force module determine the parameters in the attention module of the m+1 th sub-model, and all the m+1 sub-models can be trained during the training process to improve training efficiency. For example, when m is 3, the parameter weights in the attention module of the first sub-model, the parameter weights in the attention module of the second sub-model, and the parameters in the attention module of the third sub-model can be sequentially determined weight, and then determine the parameters in the attention module of the fourth sub-model according to the parameter weights in the attention module of the third sub-model to obtain the probability of occurrence of the fourth task, and according to the probability of occurrence of the fourth task and the corresponding Based on the sample results of the fourth task, the parameters of the four sub-models are optimized and trained.

In the embodiment of the present specification, the multi-task model may correspond to a sub-model for each task in the multi-task, and transfer the associated information of the previous task to the next task through the attention module between adjacent tasks, so that the The next task can be combined with the related information of the previous task to obtain a more accurate and more relevant prediction result.

As an option of this embodiment, the mth sub-model further includes an embedded module, a combination module, and a first conversion module;

Determine the parameter weights of the sample features in the attention module of the mth sub-model, including:

Input the sample user features and sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and sample product features respectively;

Input the feature vectors corresponding to the sample user features and the sample product features respectively into the combination module to obtain the combined feature vector;

inputting the combined feature vector into the first conversion module to obtain the first conversion information;

The first conversion information is input to the attention module of the mth sub-model, and the parameter weight of the sample feature in the mth sub-model is obtained.

Specifically, reference may be made here to a schematic structural diagram of a sub-model provided by an embodiment of the present specification shown in FIG. 4 . As shown in FIG. 4 , the schematic structural diagram can be used to represent the structural schematic diagram of the m th sub-model in this embodiment, and the m th sub-model can sequentially include an embedding module, a combining module, a first transformation module, and an attention module in order of connection. The embedding module can be used to input the sample user features and the sample product features in the sample features, and obtain feature vectors corresponding to the sample user features and the sample product features respectively. It can be understood that the sample user features here can be expressed as the identity features of the sample users, such as but not limited to including the user name, user classification, and user address, etc., and the sample product features can be expressed as the features used to represent the products corresponding to the events. Sample feature information, such as but not limited to sample product names including events, sample product production information, and sample product definition information, the sample product definition information may be understood as information used to characterize product functions. The embedding module in this embodiment can be, but is not limited to, an embedding layer (that is, an Embedding layer) in the speech processing technology. A corpus containing a dictionary and characters can be preset in the embedding module, and each character can be Corresponding to a Chinese character in the dictionary, for example, the Chinese character "you" may correspond to character 1, and the Chinese character "好" may correspond to character 3. When the sample user features and sample product features are input into the embedding module, each Chinese character in the sample user features and sample product features can be converted into corresponding characters according to the corpus, and the sample user features and The feature vector corresponding to the sample product feature is output to the combination module. It can also be understood that the order of inputting the sample user features and the sample product features is not limited here. If possible, the sample user features can be input into the embedding module to obtain the feature vector corresponding to the sample user features, and then the sample product features can be input. Input to the embedding module to obtain the feature vector corresponding to the sample product features. Possibly, the sample product features can be input into the embedding module first to obtain feature vectors corresponding to the sample product features, and then the sample user features can be input into the embedding module to obtain the feature vectors corresponding to the sample user features. Possibly, the sample user feature and the sample product feature can be input into the embedding module at the same time, and the feature vector corresponding to the sample user feature and the feature vector corresponding to the sample product feature can be obtained at the same time.

Further, after the feature vectors respectively corresponding to the sample user features and the sample product features are input to the combining module, the combining module can output the combined feature vector. Among them, the combined feature vector can be understood as the combination of the feature vector corresponding to the sample user feature and the feature vector corresponding to the sample product feature. Here, the feature vector corresponding to the sample user feature can be expressed as [A, B], and the feature vector corresponding to the sample user feature The feature vector corresponding to the product feature can be expressed as [C, D, E] as an example, and the combined feature vector obtained by the combination module can be expressed as [A, B, C, D, E].

Further, after the combined feature vector is obtained according to the combining module, the combined feature vector can be input into the first conversion module to obtain the first conversion information. The first conversion module can be understood as a neural network corresponding to the mth task, and is used to obtain vector information that can represent the corresponding features of the mth task according to the combined feature vector. It can be understood that the first conversion module can be, but is not limited to, a deep neural network (Deep Neural Networks, DNN), a deep interest learning network (Deep Interest Network, DIN) or a deep recommendation network (Factorization-Machinebased Neural Network, DeepFM). ) and any other deep learning neural network.

Further, after obtaining the first conversion information according to the first conversion module, the first conversion information can be input to the attention module, and the parameter weight of the sample feature in the mth sub-model can be obtained by learning. Wherein, the attention module can learn the vector information of the feature corresponding to the association information between the mth task and the m+1th task in the first conversion information, so as to obtain the relationship between the mth task and the m+1th task. The parameter weights corresponding to the vector information of the features corresponding to the associated information between the tasks. It can be understood that, the attention module here may be, but not limited to, an attention mechanism (ie, an attention mechanism) in a common technology, and this embodiment is not limited to this.

As another option of this embodiment, the m+1 th sub-model further includes an embedded module, a combination module, and a second conversion module;

The multi-task model is trained based on sample features, parameter weights in the attention module of the m-th sub-model, and parameters in the attention module of the m+1-th sub-model, including:

Input the sample user features and sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and sample product features respectively;

Input the feature vectors corresponding to the sample user features and the sample product features respectively into the combination module to obtain the combined feature vector;

inputting the combined feature vector into the second conversion module to obtain second conversion information;

Input the second conversion information into the attention module of the m+1 th sub-model, and obtain the prediction of the task corresponding to the m+1 th sub-model according to the second conversion information and the parameters in the attention module of the m+1 th sub-model result;

The multi-task model is trained according to the prediction result of the task corresponding to the m+1 th sub-model and the user's sample result of the task corresponding to the m+1 th sub-model.

Specifically, reference may be made here to the schematic structural diagram of a multi-task model provided by an embodiment of the present specification shown in FIG. 5 . As shown in FIG. 5 , the multi-task model may include the m-th sub-model and the m+1-th sub-model, and the m-th sub-model may correspond to the m-th task in the multi-task corresponding to the sample feature. One sub-model may correspond to the m+1 th task in the multi-task corresponding to the sample feature. Wherein, the mth sub-model may include an embedding module, a combination module, a first transformation module, and an attention module in order of connection, and the m+1th sub-model may include an embedding module, a combination module, a second transformation module, and a second transformation module in order of connection. attention module. It can be understood that the embedded module in the mth sub-model and the embedded module in the m+1th sub-model can be the same embedded module (it can also be understood as sharing the same embedded module), which can effectively retain the Association information between m tasks and the m+1th task. Similarly, the combination module in the mth sub-model can also be the same combination module as the combination module in the m+1th submodel, so as to ensure the association between the mth task and the m+1th task reliability of information.

In the m+1th sub-model, the embedding module can be used to input the sample user features and sample product features in the sample features, and obtain feature vectors corresponding to the sample user features and sample product features respectively. It can be understood that the sample user features here can be expressed as the identity features of the sample users, such as but not limited to including the user name, user classification, and user address, etc., and the sample product features can be expressed as the features used to represent the products corresponding to the events. Sample feature information, such as but not limited to sample product names including events, sample product production information, and sample product definition information, the sample product definition information may be understood as information used to characterize product functions. The embedding module in this embodiment can be, but is not limited to, an embedding layer (that is, an Embedding layer) in the speech processing technology. A corpus containing a dictionary and characters can be preset in the embedding module, and each character can be Corresponding to a Chinese character in the dictionary, for example, the Chinese character "you" may correspond to character 1, and the Chinese character "好" may correspond to character 3. When the sample user features and sample product features are input into the embedding module, each Chinese character in the sample user features and sample product features can be converted into corresponding characters according to the corpus, and the sample user features and The feature vector corresponding to the sample product feature is output to the combination module. It can also be understood that the order of inputting the sample user features and the sample product features is not limited here. If possible, the sample user features can be input into the embedding module to obtain the feature vector corresponding to the sample user features, and then the sample product features can be input. Input to the embedding module to obtain the feature vector corresponding to the sample product features. Possibly, the sample product features can be input into the embedding module first to obtain feature vectors corresponding to the sample product features, and then the sample user features can be input into the embedding module to obtain the feature vectors corresponding to the sample user features. Possibly, the sample user feature and the sample product feature can be input into the embedding module at the same time, and the feature vector corresponding to the sample user feature and the feature vector corresponding to the sample product feature can be obtained at the same time.

Further, after the feature vectors respectively corresponding to the sample user features and the sample product features are input to the combining module, the combining module can output the combined feature vector. Among them, the combined feature vector can be understood as the combination of the feature vector corresponding to the sample user feature and the feature vector corresponding to the sample product feature. Here, the feature vector corresponding to the sample user feature can be expressed as [A, B], and the feature vector corresponding to the sample user feature The feature vector corresponding to the product feature can be expressed as [C, D, E] as an example, and the combined feature vector obtained by the combination module can be expressed as [A, B, C, D, E].

Further, after the combined feature vector is obtained according to the combining module, the combined feature vector may be input into the second conversion module to obtain second conversion information. The second conversion module can be understood as a neural network corresponding to the m+1 th task, and is used to obtain vector information that can represent the corresponding features of the m+1 th task according to the combined feature vector. It can be understood that the second conversion module can be, but is not limited to, a deep neural network (Deep Neural Networks, DNN), a deep interest learning network (Deep Interest Network, DIN) or a deep recommendation network (Factorization-Machinebased Neural Network, DeepFM). ) and any other deep learning neural network.

Further, after the second conversion information is obtained according to the second conversion module, the second conversion information can be input into the attention module, and the second conversion information can be obtained according to the parameters adjusted in the attention module and the activation function sigmoid in the attention module. The prediction result of the m+1th task. It can be understood that the parameters of the attention module in the m+1 th sub-model may contain the association information with the m th task, and the m th task obtained by the attention module in the m+1 th sub-model can be obtained. The probability of occurrence of +1 task has a correlation with the mth task, which in turn makes the prediction result more reliable and accurate than common techniques.

It can also be understood that a fully connected module can be set between the attention module in the m+1th sub-model and the attention module in the m-th sub-model. When the parameters in the attention module of the m-th sub-model are determined After the weighting, the parameter weights in the attention module of the mth sub-model can be input into the fully connected module, so as to obtain connection information including the parameter weights in the attention module of the mth sub-model. It should be noted here that the fully connected module in this embodiment can be used to retain the parameter weights in the attention module of the mth sub-model and transmit the parameter weights in the attention module of the mth sub-model. The integrity of the parameter weights in the attention module of the m-th sub-model is effectively guaranteed. After the fully-connected module obtains the connection information containing the parameter weights in the attention module of the m-th sub-model, the fully-connected module can input the connection information containing the parameter weights in the attention module of the m-th sub-model to In the attention module of the m+1 th sub-model, the parameters in the attention module of the m+1 th sub-model are adjusted in combination with the parameter weights in the attention module of the m-th sub-model.

Further, after the prediction result of the task corresponding to the m+1 th sub-model is obtained, the prediction result of the task corresponding to the m+1 th sub-model can be combined with the sample result of the user for the task corresponding to the m+1 th sub-model, to many The parameters in the task model are optimized and trained until the prediction result of the task corresponding to the m+1 th sub-model is closer to the user's sample result for the task corresponding to the m+1 th sub-model. It can be understood that in this embodiment, the multi-task model can be trained by, but not limited to, the prediction result of the m+1 th sub-model and the sample result. The task model is trained, not limited to this.

It should be noted that the above-mentioned multi-task model may include, but is not limited to, only the mth sub-model and the m+1th sub-model, for example, the first sub-model, the second sub-model...the m-th sub-model may also be included. +2 sub-models, m+3 sub-models, etc., each sub-model can correspond to a task in the corresponding multi-task of the sample feature, and each sub-model can include an embedded module, a combination module, The corresponding transformation modules and attention modules. It can be understood that a fully connected module can be set between the attention modules of any two adjacent sub-models, so as to transfer the parameter weights in the attention module of the previous sub-model to the attention of the latter sub-model. force module.

As another option of this embodiment, after obtaining the prediction result of the task corresponding to the m+1 th sub-model according to the second conversion information and the parameters of the m+1 th sub-model, the method further includes:

Obtain the first loss function according to the prediction result of the task corresponding to the m+1 th sub-model and the sample characteristics;

The m+1 th sub-model is optimized based on the first loss function.

Specifically, in the process of training the multi-task model, after obtaining the prediction result of the task corresponding to the m+1th sub-model, the cross-entropy loss function (that is, corresponding to the first loss function mentioned above) can be calculated by calculating way to optimize the parameters of the m+1th sub-model. Among them, the calculation method of the cross entropy loss function is shown in the following formula (1):

可表示为第m+1个子模型对应的任务的预测结果。此处y_t的取值可以但不局限于为1或0，0可表征第m+1个任务未完成，1可表征第m+1个任务已完成。In formula (1), L _ce can be expressed as the cross-entropy loss function of the m+1th sub-model, θ is the parameter of the m+1th model, N is the number of sample features, D is the sample set, (x, y _t ) is a set of samples in the sample set D, x can be represented as the sample user feature and sample product feature corresponding to the m+1th task in the sample, y _t can be represented as the user’s sample corresponding to the m+1th task result,

It can be expressed as the prediction result of the task corresponding to the m+1th sub-model. The value of y _t here may be, but not limited to, 1 or 0, 0 may indicate that the m+1 th task has not been completed, and 1 may indicate that the m+1 th task has been completed.

As another option of this embodiment, after the first loss function is obtained according to the prediction result of the task corresponding to the m+1 th sub-model and the sample characteristics, before the m+1 th sub-model is optimized based on the first loss function ,Also includes:

Determine whether the prediction result of the task corresponding to the m+1th sub-model satisfies the preset condition;

When it is determined that the prediction result of the task corresponding to the m+1 th sub-model does not meet the preset condition, obtain the prediction result of the task corresponding to the m th sub-model according to the first conversion information and the parameters in the attention module of the m th sub-model;

Obtain the second loss function according to the prediction result of the task corresponding to the mth submodel, the prediction result of the task corresponding to the m+1th submodel, and the sample characteristics;

Optimize the m+1 th sub-model based on the first loss function, including:

The m+1 th sub-model is optimized based on the first loss function and the second loss function.

According to the actual situation, only when the mth task is completed, it is possible to complete the m+1th task, that is to say, the prediction result corresponding to the mth task is higher than the prediction result corresponding to the m+1th task. . Based on this, it is possible to first judge whether the prediction result corresponding to the m+1 th task meets the requirements, and then perform parameter optimization of the m+1 th sub-model according to the judgment result.

Specifically, after calculating the cross-entropy loss function of the m+1 th sub-model, the prediction result of the task corresponding to the m+1 th sub-model can be compared with the preset result, and after determining the m+1 th sub-model When the prediction result of the corresponding task is higher than the preset result, first obtain the prediction result of the m-th task according to the parameters in the attention module of the m-th sub-model and the activation function sigmoid in the attention module, and then combine The prediction result of the task corresponding to the m th sub-model, the prediction result of the task corresponding to the m+1 th sub-model, and the sample features are obtained to obtain a probability calibration loss function (that is, the second loss function mentioned above). Among them, before the prediction result of the m-th task is obtained, the parameters in the attention module of the m-th sub-model can be determined according to the parameter weights in the attention module of the m-1-th sub-model, and then the sample user characteristics and sample The product features are input into the mth sub-model, and then go through the embedding module, the combination module, the first conversion module and the attention module in turn, and the prediction result of the mth task is obtained through the activation function sigmoid in the attention module.

It can be understood that the above-mentioned preset result can be determined according to the sample result corresponding to the mth task of the user, for example, but not limited to, the preset result can be set to be smaller than the sample result corresponding to the user's mth task.

The calculation method of the probability calibration loss function here can be obtained by the following formula (2):

可表示为第m个子模型对应的任务的预测结果，

可表示为第m+1个子模型对应的任务的预测结果。此处当

大于

时(也即是说第m+1个子模型对应的任务的预测结果大于第m个子模型对应的任务的预测结果)，表明第m+1个子模型对应的任务的预测结果不符合实际情况，需要引入该概率校准损失函数对多任务模型进行校准。In formula (2), L _le can be expressed as the probability calibration loss function of the m+1th sub-model,

can be expressed as the prediction result of the task corresponding to the mth sub-model,

It can be expressed as the prediction result of the task corresponding to the m+1th sub-model. here when

more than the

(that is to say, the prediction result of the task corresponding to the m+1th submodel is greater than the prediction result of the task corresponding to the mth submodel), it indicates that the prediction result of the task corresponding to the m+1th submodel does not conform to the actual situation. The probabilistic calibration loss function is introduced to calibrate the multi-task model.

For ease of understanding, when optimizing the m+1 th sub-model based on the first loss function and the second loss function, the target of the m+1 th sub-model can be determined first according to the first loss function and the second loss function. The loss function is used to optimize the parameters of the m+1th sub-model through the target loss function. Here, the objective loss function of the m+1th sub-model can be calculated by the following formula (3):

L(θ)=L _ce (θ)+αL _le (θ) (3)

In formula (3), L(θ) can be expressed as the objective loss function of the m+1th sub-model, L _ce (θ) can be expressed as the cross-entropy loss function of the m+1th sub-model, and L _le (θ) can be expressed as is the probability calibration loss function of the m+1 th sub-model, α can be expressed as a calibration parameter, and the larger α is, the greater the weight of the probability calibration loss function of the m+1 th sub-model.

It can be understood that, the above formula (3) can also perform parameter optimization for any other sub-model in the multi-task model, and this embodiment is not limited to this.

It should be noted that, in order to further improve the parameter optimization effect of the m+1 th sub-model of the multi-task model, the mean square error module can also be added to the m th sub-model and the m+1 th sub-model respectively. And the sample product feature can be abstracted into a mapping function when it passes through the embedding module, and the mean square error loss function of the m+1 th sub-model is constructed based on the mapping function of the m-th sub-model and the mapping function of the m+1-th sub-model respectively. , and combine the above-mentioned cross entropy loss function, probability calibration loss function and the mean square error loss function to optimize the parameters of the m+1th sub-model.

Among them, the mean square error loss function of the m+1th sub-model can be calculated by the following formula (4):

In formula (4), L _mse can be expressed as the mean square error loss function of the m+1 th sub-model, f _t ( _xi ) can be expressed as the mapping function of the m+1 th sub-model, f _t-1 (x _i ) can be expressed as the mapping function of the mth submodel.

Combining the above formula (3) and the mean square error loss function, the target loss function of the new m+1th sub-model can be obtained by the following formula (5):

L(θ)=L _ce (θ)+αL _le (θ)+γL _mse (5)

In formula (5), L(θ) can be expressed as the objective loss function of the m+1th sub-model, L _ce (θ) can be expressed as the cross-entropy loss function of the m+1th sub-model, and L _le (θ) can be expressed as is the probability calibration loss function of the m+1 th sub-model, L _mse can be expressed as the mean square error loss function of the m+1 th sub-model, α can be expressed as a calibration parameter, γ can be expressed as an error parameter, and γ can be any Natural number.

As another option of this embodiment, after obtaining the prediction result of the task corresponding to the m th sub-model according to the first conversion information and the parameter weight of the m th sub-model, the method further includes:

According to the prediction result of the task corresponding to the mth sub-model and the sample characteristics, the third loss function is obtained;

The m-th sub-model is optimized based on the third loss function.

In the process of training the multi-task model, in addition to optimizing the parameters of the m+1th sub-model, parameters of any sub-model in the multi-task model can also be optimized to further ensure the accuracy of the prediction results of the multi-task model. sex.

Specifically, after obtaining the prediction result of the task corresponding to the m-th sub-model, the parameters of the m-th sub-model can be optimized by calculating the cross-entropy loss function (that is, corresponding to the third loss function mentioned above). The calculation method of the cross-entropy loss function can refer to the above formula (1), which is not repeated here.

Referring to FIG. 6, FIG. 6 shows a schematic flowchart of a multi-task prediction method provided by an embodiment of the present specification.

As shown in Figure 6, the multi-task prediction method may at least include the following steps:

Step 602: Determine the parameter weight of the target feature in the attention module of the mth sub-model.

Specifically, the target features may include target user features and target product features, where the target user features can be understood as target feature information of the user who wants to execute the event, such as the user's target identity information, which may specifically include user name, user classification or user address and other at least one kind of information. The target product feature can be understood as the target feature information used to represent the product corresponding to the event, such as the event target product name, target product production information, or target product definition information and any other at least one kind of information, the target product definition information can be understood Information used to characterize product functionality.

After the target user characteristics and the target product characteristics are first determined, the parameter weights in the attention module of the model corresponding to the mth task in the multi-task can be obtained according to the target user characteristics and the target product characteristics. The multi-task model in this embodiment may include M sub-models, each sub-model may include an attention module, each task in the multi-task corresponding to the target feature may correspond to a sub-model, and the adjacent tasks can correspond to two adjacent sub-models. For example, the first task in the multi-task corresponding to the target feature may correspond to the first sub-model in the multi-task model, and the second task in the multi-task corresponding to the target feature may correspond to the second sub-model in the multi-task model sub-model, the m-th task in the multi-task corresponding to the target feature can correspond to the m-th sub-model in the multi-task model, that is, the attention of the model corresponding to the m-th task in the multi-task mentioned above The parameter weight in the module can be understood as the parameter weight in the attention module of the mth sub-model in the multi-task model. Here, M may be a positive integer greater than m, for example, when m is 2, M may be a positive integer greater than 2.

It can be understood that the m-th task in the multi-task corresponding to the target features mentioned above can correspond to the m-th sub-model in the multi-task model, and the parameter weights in the attention module in the m-th sub-model can be used to represent The parameter corresponding to the association information between the m th task and the m+1 th task may also correspond to some parameters of all parameters in the attention module of the sub-model corresponding to the m th task. Here, all parameters in the attention module of the sub-model corresponding to the m-th task can be represented as A, B, C, D and E as an example, the parameter weights in the attention module of the sub-model corresponding to the m-th task It can be expressed as (0, 1, 1, 0, 0) but is not limited to, that is, the parameters corresponding to the association information between the m th task and the m+1 th task include B and C.

It can also be understood that the task corresponding to the mth sub-model in this embodiment may be any one of the tasks from the first task to the penultimate task in the multi-task corresponding to the target feature. Possibly, when the number of multi-tasks corresponding to the target feature is 5, the value of m may be 1 or 2 or 3 or 4. Possibly, when the number of multi-tasks corresponding to the target feature is 2, the value of m can only be 1.

Step 604: Determine the parameters in the attention module of the m+1 th sub-model according to the parameter weight in the attention module of the m th sub-model.

Specifically, after the parameter weight in the attention module of the mth sub-model is determined, the parameter weight in the attention module of the m-th sub-model can be transferred to the attention module of the adjacent m+1-th sub-model , to adjust the parameters in the attention module of the m+1th sub-model. Here, the parameters in the attention module of the m+1th model can be expressed as B, D, E and F, and the parameters corresponding to the parameter weights in the attention module of the mth model can be expressed as B and C as an example. After adjustment The parameters in the attention module of the m+1 th model can be denoted as B, C, D, E, and F.

It can be understood that the attention modules included in each sub-model in the multi-task model of this embodiment may adopt the same structure, and the difference lies in that the parameters of the attention modules included in each sub-model are different.

Step 606 , based on the target feature and the parameters in the attention module of the m+1 th sub-model, obtain the prediction result of the task corresponding to the m+1 th sub-model.

In this embodiment, the mth sub-model may include an embedding module, a combination module, a first transformation module, and an attention module in the order of connection, and the m+1th sub-model may include an embedding module, a combination module, and a second transformation in the order of connection. module and attention module. It can be understood that the embedded module in the mth sub-model and the embedded module in the m+1th sub-model can be the same embedded module (it can also be understood as sharing the same embedded module), which can effectively retain the Association information between m tasks and the m+1th task. Similarly, the combination module in the mth sub-model can also be the same combination module as the combination module in the m+1th submodel, so as to ensure the association between the mth task and the m+1th task reliability of information.

Specifically, in the m+1 th sub-model, the embedding module can be used to input the target user features and the target product features in the target features, and obtain feature vectors corresponding to the target user features and the target product features respectively. It can be understood that the target user feature here can be expressed as the identity feature of the target user, such as but not limited to including user name, user classification and user address, etc., and the target product feature can be expressed as a feature used to represent the product corresponding to the event. The target feature information includes, for example, but not limited to, the name of the target product of the event, the production information of the target product, and the target product definition information. The target product definition information may be understood as information used to characterize the function of the product. The embedding module in this embodiment can be, but is not limited to, an embedding layer (that is, an Embedding layer) in the speech processing technology. A corpus containing a dictionary and characters can be preset in the embedding module, and each character can be Corresponding to a Chinese character in the dictionary, for example, the Chinese character "you" may correspond to character 1, and the Chinese character "好" may correspond to character 3. When the target user feature and the target product feature are input into the embedded module, each Chinese character in the target user feature and the target product feature can be converted into a corresponding character according to the corpus, and the target user feature and The feature vector corresponding to the target product feature is output to the combination module. It can also be understood that the order of inputting the target user features and the target product features is not limited here. If possible, the target user features can be input into the embedding module to obtain the feature vector corresponding to the target user features, and then the target product features can be input. Input to the embedding module to obtain the feature vector corresponding to the target product feature. Possibly, the feature vector of the target product may be input into the embedding module to obtain the feature vector corresponding to the feature of the target product, and then the feature vector corresponding to the feature of the target user may be obtained by inputting the feature of the target user into the embedding module. Possibly, the target user feature and the target product feature can be input into the embedding module at the same time, and the feature vector corresponding to the target user feature and the feature vector corresponding to the target product feature can be obtained at the same time.

Further, after the feature vectors respectively corresponding to the target user feature and the target product feature are input to the combining module, the combined feature vector can be output by the combining module. Among them, the combined feature vector can be understood as the combination of the feature vector corresponding to the target user feature and the feature vector corresponding to the target product feature. Here, the feature vector corresponding to the target user feature can be expressed as [A, B], and the feature vector corresponding to the target user feature The feature vector corresponding to the product feature can be expressed as [C, D, E] as an example, and the combined feature vector obtained by the combination module can be expressed as [A, B, C, D, E].

Further, after the combined feature vector is obtained according to the combining module, the combined feature vector may be input into the second conversion module to obtain second conversion information. The second conversion module can be understood as a neural network corresponding to the m+1 th task, and is used to obtain vector information that can represent the corresponding features of the m+1 th task according to the combined feature vector. It can be understood that the second conversion module can be, but is not limited to, a deep neural network (Deep Neural Networks, DNN), a deep interest learning network (Deep Interest Network, DIN) or a deep recommendation network (Factorization-Machinebased Neural Network, DeepFM). ) and any other deep learning neural network.

Further, after the second conversion information is obtained according to the second conversion module, the second conversion information can be input into the attention module, and the second conversion information can be obtained according to the parameters adjusted in the attention module and the activation function sigmoid in the attention module. The prediction result of the m+1th task. It can be understood that the parameters of the attention module in the m+1 th sub-model may contain the association information with the m th task, and the m th task obtained by the attention module in the m+1 th sub-model can be obtained. The probability of occurrence of +1 task has a correlation with the mth task, which in turn makes the prediction result more reliable and accurate than common techniques.

Referring to FIG. 7 , FIG. 7 shows a schematic structural diagram of a multi-task model training apparatus provided by an embodiment of the present specification.

The multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module. As shown in FIG. 7 , the multi-task model training apparatus 700 may further include at least a first processing module 701, a second processing module 702 and a training module 703, wherein:

The first processing module 701 is used to determine the parameter weight of the sample feature in the attention module of the mth sub-model; wherein, m is a positive integer smaller than M;

The second processing module 702 is configured to determine parameters in the attention module of the m+1 th sub-model according to the parameter weights in the attention module of the m th sub-model; wherein the sample features include sample user features, sample Product features and user sample results for tasks corresponding to the m+1th sub-model;

A training module 703, configured to train the multi-task model based on the sample features, the parameter weights in the attention module of the m th sub-model, and the parameters in the attention module of the m+1 th sub-model .

In some possible embodiments, the mth sub-model further includes an embedding module, a combining module and a first converting module;

The first processing module 701 includes:

a first embedding unit, configured to input the sample user features and the sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and the sample product features respectively;

The first combining unit is used to input the feature vectors corresponding to the sample user features and the sample product features respectively into the combining module to obtain the combined feature vector;

a first conversion unit, configured to input the combined feature vector into a first conversion module to obtain first conversion information;

The first generating unit is configured to input the first conversion information to the attention module of the mth sub-model, and obtain the parameter weight of the sample feature in the mth sub-model.

In some possible embodiments, a fully connected module is set between the attention module of the mth sub-model and the attention module of the m+1th sub-model;

The second processing module 702 includes:

The connection unit is used for inputting the parameter weights in the attention module of the mth sub-model to the fully connected module to obtain connection information including the parameter weights in the attention module of the mth sub-model;

The second generating unit is configured to input the connection information into the attention module of the m+1 th sub-model, and obtain the parameters in the attention module of the m+1 th sub-model.

In some possible embodiments, the m+1 th sub-model further includes an embedding module, a combining module and a second converting module;

Training module 703 includes:

The second embedding unit is used for inputting the sample user features and the sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and the sample product features respectively;

The second combining unit is used to input the feature vectors corresponding to the sample user features and the sample product features respectively into the combining module to obtain the combined feature vector;

a second conversion unit, configured to input the combined feature vector into a second conversion module to obtain second conversion information;

The third generation unit is configured to input the second conversion information into the attention module of the m+1 th sub-model, and obtain the m+1 th according to the second conversion information and the parameters in the attention module of the m+1 th sub-model The prediction results of the tasks corresponding to the sub-models;

The training unit is configured to train the multi-task model according to the prediction result of the task corresponding to the m+1 th sub-model and the user's sample result of the task corresponding to the m+1 th sub-model.

In some possible embodiments, the training module 703 further includes:

The first computing unit is used to obtain the prediction result of the task corresponding to the m+1 th sub-model according to the second conversion information and the parameters of the m+1 th sub-model, according to the prediction result of the task corresponding to the m+1 th sub-model and sample features to obtain the first loss function;

The first optimization unit is configured to optimize the m+1 th sub-model based on the first loss function.

In some possible embodiments, the training module 703 further includes:

The judgment unit is used to judge the m+ Whether the prediction result of the task corresponding to a sub-model satisfies the preset condition;

The second calculation unit is configured to obtain the mth submodel according to the first conversion information and the parameters in the attention module of the mth submodel when it is determined that the prediction result of the task corresponding to the m+1th submodel does not meet the preset condition The prediction result of the corresponding task;

The third computing unit is used to obtain the second loss function according to the prediction result of the task corresponding to the mth submodel, the prediction result of the task corresponding to the m+1th submodel, and the sample characteristics;

The first optimization unit is specifically used for:

The m+1 th sub-model is optimized based on the first loss function and the second loss function.

In some possible embodiments, the training module 703 further includes:

The fourth calculation unit is used to obtain the prediction result of the task corresponding to the m th sub-model according to the first conversion information and the parameter weight of the m th sub-model, according to the prediction result of the task corresponding to the m th sub-model and the sample feature, to obtain The third loss function;

The second optimization unit is configured to optimize the mth sub-model based on the third loss function.

Referring to FIG. 8 , FIG. 8 shows a schematic structural diagram of a multi-task prediction apparatus provided by an embodiment of the present specification.

As shown in FIG. 8 , the multi-task prediction device 800 is applied to a multi-task model. The multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module. The multi-task prediction device 800 It may at least include a third processing module 801, a fourth processing module 802 and a prediction module 803, wherein:

The third processing module 801 is used to determine the parameter weight of the target feature in the attention module of the mth sub-model; wherein, m is a positive integer smaller than M;

The fourth processing module 802 is used to determine the parameters in the attention module of the m+1 th sub-model according to the parameter weight in the attention module of the m th sub-model; wherein, the target features include target user features and target product features;

The prediction module 803 is configured to obtain the prediction result of the task corresponding to the m+1 th sub-model based on the target feature and the parameters in the attention module of the m+1 th sub-model.

Please refer to FIG. 9. FIG. 9 shows a schematic structural diagram of another multi-task model training apparatus provided by an embodiment of the present specification.

The multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module. As shown in FIG. 9 , the multi-task model training apparatus 900 may further include: at least one processor 901 , at least one network interface 904 , user interface 903 , memory 905 and at least one communication bus 902 .

Wherein, the communication bus 902 can be used to realize the connection and communication of the above components.

The user interface 903 may include buttons, and the optional user interface may also include a standard wired interface and a wireless interface.

The network interface 904 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, and the like.

The processor 901 may include one or more processing cores. The processor 901 uses various interfaces and lines to connect various parts of the entire multitasking model training device 900, and by running or executing the instructions, programs, code sets or instruction sets stored in the memory 905, and calling the data, perform various functions of the multi-task model training apparatus 900 and process data. Optionally, the processor 901 may be implemented in at least one hardware form among DSP, FPGA, and PLA. The processor 901 may integrate one or a combination of a CPU, a GPU, a modem, and the like. Among them, the CPU mainly handles the operating system, user interface, and application programs; the GPU is used to render and draw the content that needs to be displayed on the display screen; the modem is used to handle wireless communication. It can be understood that, the above-mentioned modem may not be integrated into the processor 901, and is implemented by a single chip.

The memory 905 may include RAM or ROM. Optionally, the memory 905 includes non-transitory computer-readable media. Memory 905 may be used to store instructions, programs, codes, sets of codes, or sets of instructions. The memory 905 may include a stored program area and a stored data area, wherein the stored program area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), Instructions and the like used to implement the above method embodiments; the storage data area may store the data and the like involved in the above method embodiments. The memory 905 can optionally also be at least one storage device located away from the aforementioned processor 901 . As shown in FIG. 9 , the memory 905 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a multi-task model training application program.

Specifically, the processor 901 can be used to call the multi-task model training application program stored in the memory 905, and specifically perform the following operations:

Determine the parameter weight of the sample feature in the attention module of the mth sub-model; where m is a positive integer less than M;

The parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model; wherein, the sample features include sample user features, sample product features, and the user's correspondence to the m+1 th sub-model. sample results of the task;

The multi-task model is trained based on the sample features, the parameter weights in the attention module of the mth submodel, and the parameters in the attention module of the m+1th submodel.

In some possible embodiments, the mth sub-model further includes an embedding module, a combining module and a first converting module;

When the processor 901 determines the parameter weight of the sample feature in the attention module of the mth sub-model, it specifically executes:

Input the sample user features and sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and sample product features respectively;

Input the feature vectors corresponding to the sample user features and the sample product features respectively into the combination module to obtain the combined feature vector;

inputting the combined feature vector into the first conversion module to obtain the first conversion information;

The first conversion information is input to the attention module of the mth sub-model, and the parameter weight of the sample feature in the mth sub-model is obtained.

In some possible embodiments, a fully connected module is set between the attention module of the mth sub-model and the attention module of the m+1th sub-model;

When the processor 901 determines the parameters in the attention module of the m+1 th sub-model according to the parameter weights in the attention module of the m th sub-model, it specifically executes:

Input the parameter weights in the attention module of the mth sub-model to the fully connected module, and obtain the connection information including the parameter weights in the attention module of the mth sub-model;

Input the connection information to the attention module of the m+1 th sub-model, and get the parameters in the attention module of the m+1 th sub-model.

In some possible embodiments, the m+1 th sub-model further includes an embedding module, a combining module and a second converting module;

When the processor 901 trains the multi-task model based on the sample features, the parameter weights in the attention module of the m th sub-model, and the parameters in the attention module of the m+1 th sub-model, it specifically executes:

Input the sample user features and sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and sample product features respectively;

Input the feature vectors corresponding to the sample user features and the sample product features respectively into the combination module to obtain the combined feature vector;

inputting the combined feature vector into the second conversion module to obtain second conversion information;

Input the second conversion information into the attention module of the m+1 th sub-model, and obtain the prediction of the task corresponding to the m+1 th sub-model according to the second conversion information and the parameters in the attention module of the m+1 th sub-model result;

The multi-task model is trained according to the prediction result of the task corresponding to the m+1 th sub-model and the user's sample result of the task corresponding to the m+1 th sub-model.

In some possible embodiments, after obtaining the prediction result of the task corresponding to the m+1 th sub-model according to the second conversion information and the parameters of the m+1 th sub-model, the processor 901 is further configured to execute:

Obtain the first loss function according to the prediction result of the task corresponding to the m+1 th sub-model and the sample characteristics;

The m+1 th sub-model is optimized based on the first loss function.

In some possible embodiments, after the processor 901 obtains the first loss function according to the prediction result of the task corresponding to the m+1 th sub-model and the sample characteristics, the m+1 th sub-model performs the calculation on the m+1 th sub-model based on the first loss function Before optimization, also used to perform:

Determine whether the prediction result of the task corresponding to the m+1th sub-model satisfies the preset condition;

When it is determined that the prediction result of the task corresponding to the m+1 th sub-model does not meet the preset condition, obtain the prediction result of the task corresponding to the m th sub-model according to the first conversion information and the parameters in the attention module of the m th sub-model;

Obtain the second loss function according to the prediction result of the task corresponding to the mth submodel, the prediction result of the task corresponding to the m+1th submodel, and the sample characteristics;

Optimize the m+1 th sub-model based on the first loss function, including:

The m+1 th sub-model is optimized based on the first loss function and the second loss function.

In some possible embodiments, after the processor 901 obtains the prediction result of the task corresponding to the m-th sub-model according to the first conversion information and the weight parameter of the m-th sub-model, the processor 901 is further configured to execute:

According to the prediction result of the task corresponding to the mth sub-model and the sample characteristics, the third loss function is obtained;

The m-th sub-model is optimized based on the third loss function.

Please refer to FIG. 10. FIG. 10 shows a schematic structural diagram of another multi-task prediction apparatus provided by an embodiment of the present specification.

The multi-task prediction apparatus 1000 is applied to a multi-task model. The multi-task model includes M sub-models, each sub-model corresponds to a task, and each sub-model includes an attention module. As shown in FIG. 10 , the multi-task prediction apparatus 1000 may further include: at least one processor 1001 , at least one network interface 1004 , user interface 1003 , memory 1005 and at least one communication bus 1002 .

Wherein, the communication bus 1002 can be used to realize the connection and communication of the above components.

The user interface 1003 may include buttons, and the optional user interface may also include a standard wired interface and a wireless interface.

Wherein, the network interface 1004 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, and the like.

The processor 1001 may include one or more processing cores. The processor 1001 uses various interfaces and lines to connect various parts of the entire multitasking prediction apparatus 1000, and by running or executing the instructions, programs, code sets or instruction sets stored in the memory 1005, and calling the data stored in the memory 1005. , which executes various functions of the multitasking prediction apparatus 1000 and processes data. Optionally, the processor 1001 may be implemented in at least one hardware form among DSP, FPGA, and PLA. The processor 1001 may integrate one or a combination of a CPU, a GPU, a modem, and the like. Among them, the CPU mainly handles the operating system, user interface, and application programs; the GPU is used to render and draw the content that needs to be displayed on the display screen; the modem is used to handle wireless communication. It can be understood that, the above-mentioned modem may not be integrated into the processor 1001, but is implemented by a single chip.

The memory 1005 may include RAM or ROM. Optionally, the memory 1005 includes non-transitory computer-readable media. Memory 1005 may be used to store instructions, programs, codes, sets of codes, or sets of instructions. The memory 1005 may include a stored program area and a stored data area, wherein the stored program area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), Instructions and the like used to implement the above method embodiments; the storage data area may store the data and the like involved in the above method embodiments. Optionally, the memory 1005 may also be at least one storage device located away from the aforementioned processor 1001 . As shown in FIG. 10 , the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module and a multitasking prediction application program.

Specifically, the processor 1001 can be used to call the multitasking prediction application program stored in the memory 1005, and specifically perform the following operations:

Determine the parameter weight of the target feature in the attention module of the mth sub-model; where m is a positive integer less than M;

The parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model; wherein, the target features include target user features and target product features;

Based on the target feature and the parameters in the attention module of the m+1 th sub-model, the prediction result of the task corresponding to the m+1 th sub-model is obtained.

The embodiments of the present specification also provide a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, when the computer or the processor is executed, the computer or the processor causes the computer or the processor to execute the steps shown in FIG. 3 or FIG. 6 above. one or more steps in an example embodiment. If each component module of the above electronic device is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in the computer-readable storage medium.

In the above-mentioned embodiments, it may be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented in software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present specification are generated. The computer may be a general purpose computer, special purpose computer, computer network, or other programmable device. The computer instructions may be stored in or transmitted over a computer-readable storage medium. The computer instructions can be sent from one website site, computer, server, or data center to another by wire (eg, coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.) A website site, computer, server or data center for transmission. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes an integration of one or more available media. The available media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, Digital Versatile Disc (DVD)), or semiconductor media (eg, Solid State Disk (SSD) ))Wait.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, The processes of the embodiments of the various methods described above may be included. The aforementioned storage medium includes various media that can store program codes, such as ROM, RAM, magnetic disk, or optical disk. The technical features in this embodiment and the implementation can be combined arbitrarily if there is no conflict.

The above-mentioned embodiments are only the preferred embodiments of this specification to describe, and do not limit the scope of this specification. Without departing from the design spirit of this specification, those of ordinary skill in the art can make technical solutions of this specification. Various modifications and improvements shall fall within the protection scope determined by the claims of this specification.

Claims (13)

1. A multi-task model training method, the multi-task model comprises M sub-models, each of the sub-models corresponds to a task respectively, and each of the sub-models comprises an attention module respectively, and the method comprises: Determine the parameter weight of the sample feature in the attention module of the mth sub-model; where m is a positive integer less than M; The parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model; wherein, the sample features include sample user features, sample product features, and the user’s preference for the m+ th sub-model. Sample results of tasks corresponding to 1 sub-model; The multi-task model is trained based on the sample features, parameter weights in the attention module of the m-th sub-model, and parameters in the attention module of the m+1-th sub-model. 2. The method according to claim 1, the m-th sub-model further comprises an embedding module, a combining module and a first converting module; The determining the parameter weight of the sample feature in the attention module of the mth sub-model includes: Inputting the sample user features and the sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and the sample product features respectively; Inputting the feature vectors corresponding to the sample user features and the sample product features respectively into the combining module to obtain a combined feature vector; Inputting the combined feature vector to the first conversion module to obtain first conversion information; The first conversion information is input into the attention module of the m-th sub-model, and the parameter weight of the sample feature in the m-th sub-model is obtained. 3. The method according to claim 1, a fully connected module is provided between the attention module of the m th sub-model and the attention module of the m+1 th sub-model; The parameters in the attention module of the m+1 th sub-model are determined according to the parameter weights in the attention module of the m th sub-model, including: inputting the parameter weights in the attention module of the mth sub-model to the fully connected module, to obtain connection information including the parameter weights in the attention module of the mth sub-model; The connection information is input into the attention module of the m+1 th sub-model to obtain the parameters in the attention module of the m+1 th sub-model. 4. The method according to claim 2, wherein the m+1 th sub-model further comprises the embedding module, the combining module and the second converting module; The multi-task model is trained based on the sample features, the parameter weights in the attention module of the m th sub-model, and the parameters in the attention module of the m+1 th sub-model, including: Inputting the sample user features and the sample product features into the embedding module to obtain feature vectors corresponding to the sample user features and the sample product features respectively; Inputting the feature vectors respectively corresponding to the sample user features and the sample product features into the combination module to obtain the combined feature vector; Inputting the combined feature vector to the second conversion module to obtain second conversion information; Inputting the second conversion information into the attention module of the m+1th sub-model, and obtaining the second conversion information and parameters in the attention module of the m+1th sub-model The prediction results of tasks corresponding to m+1 sub-models; The multi-task model is trained according to the prediction result of the task corresponding to the m+1 th sub-model and the user's sample result of the task corresponding to the m+1 th sub-model. 5. The method according to claim 4, wherein after obtaining the prediction result of the task corresponding to the m+1 th sub-model according to the second conversion information and the parameters of the m+1 th sub-model, the Methods also include: Obtain a first loss function according to the prediction result of the task corresponding to the m+1 th sub-model and the sample feature; The m+1 th sub-model is optimized based on the first loss function. 6. The method according to claim 5, wherein after the first loss function is obtained according to the prediction result of the task corresponding to the m+1 th sub-model and the sample feature, the first loss function is obtained based on the first loss function. Before optimizing the m+1 th sub-model, the method further includes: Judging whether the prediction result of the task corresponding to the m+1 th sub-model satisfies a preset condition; When it is determined that the prediction result of the task corresponding to the m+1 th sub-model does not meet the preset condition, obtain the The prediction results of the tasks corresponding to the m sub-models; According to the prediction result of the task corresponding to the m th sub-model, the prediction result of the task corresponding to the m+1 th sub-model, and the sample feature, a second loss function is obtained; The optimizing the m+1 th sub-model based on the first loss function includes: The m+1 th sub-model is optimized based on the first loss function and the second loss function. 7. The method according to claim 6, after obtaining the prediction result of the task corresponding to the m-th sub-model according to the first conversion information and the weight parameter of the m-th sub-model, the method further comprises: : According to the prediction result of the task corresponding to the m th sub-model and the sample feature, a third loss function is obtained; The m-th sub-model is optimized based on the third loss function. 8. A multi-task prediction method, the method is applied to a multi-task model, the multi-task model includes M sub-models, each of the sub-models corresponds to a task respectively, and each of the sub-models includes an attention module, the method includes: Determine the parameter weight of the target feature in the attention module of the mth sub-model; where m is a positive integer less than M; Determine the parameters in the attention module of the m+1 th sub-model according to the parameter weights in the attention module of the m th sub-model; wherein, the target features include target user features and target product features; Based on the target feature and the parameters in the attention module of the m+1 th sub-model, the prediction result of the task corresponding to the m+1 th sub-model is obtained. 9. A multi-task model training device, the multi-task model comprises M sub-models, each of the sub-models corresponds to a task respectively, and each of the sub-models comprises an attention module respectively, and the The training device includes: The first processing module is used to determine the parameter weight of the sample feature in the attention module of the mth sub-model; wherein, m is a positive integer smaller than M; The second processing module is configured to determine the parameters in the attention module of the m+1 th sub-model according to the parameter weight in the attention module of the m th sub-model; wherein, the sample features include sample user features, sample products Features and the user's sample results for the task corresponding to the m+1th sub-model; A training module, configured to train the multi-task model based on the sample features, parameter weights in the attention module of the mth sub-model, and parameters in the attention module of the m+1th sub-model. 10. A multi-task prediction device, the device is applied to a multi-task model, the multi-task model includes M sub-models, each of the sub-models corresponds to a task, and each of the sub-models includes an attention module, the device includes: The third processing module is used to determine the parameter weight of the target feature in the attention module of the mth sub-model; wherein, m is a positive integer smaller than M; The fourth processing module is used to determine the parameters in the attention module of the m+1 th sub-model according to the parameter weight in the attention module of the m th sub-model; wherein, the target features include target user features and target products feature; A prediction module, configured to obtain the prediction result of the task corresponding to the m+1 th sub-model based on the target feature and the parameters in the attention module of the m+1 th sub-model. 11. A multi-task model training device, comprising a processor and a memory; the processor is connected to the memory; the memory for storing executable program codes; The processor executes a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute the method according to any one of claims 1-7. 12. A multi-task prediction device, comprising a processor and a memory; the processor is connected to the memory; the memory for storing executable program codes; The processor executes a program corresponding to the executable program code by reading the executable program code stored in the memory for performing the method of claim 8 . 13. A computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the method according to any one of claims 1-8 is implemented.

Images