CN112070162A - Multi-class processing task training sample construction method, device and medium - Google Patents

Abstract

The invention discloses a multi-class processing task training sample construction method, equipment and medium. The method includes: acquiring the real probability distribution of sample data on multiple preset categories, and determining its predicted probability distribution; according to the real probability distribution and Predict the probability distribution, and determine the loss list of the sample data on multiple preset categories; determine the mask list according to the loss list and the true probability distribution, and determine the forward category loss and multi-level loss of the sample data according to the mask list and the loss list. A negative category loss; then the two determine the positive category and multiple negative categories that the sample data belongs to in multiple preset categories, and according to the positive category and multiple negative categories, respectively construct multiple presets Positive and negative samples of the class. The present invention constructs positive and negative samples on multiple preset categories from the positive category loss and multiple negative category losses of multiple pieces of sample data, which is beneficial to the balance of sample data among each preset category.

Description

Method, equipment and medium for constructing training samples for multi-class processing tasks

technical field

The present invention relates to the technical field of financial technology (Fintech), and in particular, to a method, device and medium for constructing training samples for multi-category processing tasks.

Background technique

With the continuous development of financial technology (Fintech), especially Internet technology finance, more and more technologies (such as artificial intelligence, big data, cloud storage, etc.) are applied in the financial field, but the financial field has also proposed various technologies. Higher requirements, such as requiring more balanced sample data in artificial intelligence.

At present, for multi-category classification tasks, the collected data of each category is directly used as sample data for multi-category classification tasks, but the amount of data collected by different categories is usually difficult to balance. A smaller number, and a larger number for popular categories. As a result, when performing multi-class classification tasks based on unbalanced samples, the accuracy of the class with a small number of samples is low.

Therefore, in the multi-class classification task, how to construct a balanced sample to ensure the accuracy of the multi-class classification task is a technical problem that needs to be solved urgently.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a multi-class processing task training sample construction method, device and medium, aiming to solve the technical problem of how to construct balanced samples in multi-class classification tasks in the prior art.

To achieve the above object, the present invention provides a method for constructing training samples for multi-category processing tasks, and the method for constructing training samples for multi-category processing tasks includes the following steps:

Obtain the real probability distribution of the sample data on the multiple preset categories, and determine the predicted probability distribution of the sample data on the multiple preset categories;

According to the real probability distribution and the predicted probability distribution, determine the loss list of the sample data in multiple preset categories;

Determine a mask list according to the loss list and the true probability distribution, and determine a positive category loss and a plurality of negative category losses of the sample data according to the mask list and the loss list;

According to the positive category loss and a plurality of the negative category losses, determine a positive category and a plurality of negative categories to which the sample data belongs in a plurality of preset categories, and according to the positive category and the multiple negative categories A multi-category classification model is generated based on the positive and negative samples of the multiple preset categories for category classification.

Optionally, after the step of respectively constructing positive and negative samples of multiple preset categories according to the positive category and multiple negative categories, the method further includes:

training a preset multi-category model based on a plurality of positive and negative item samples of the preset categories to generate a multi-category classification model;

When the data to be classified is received, the data to be classified is classified based on the multi-class classification model, and the class to which the data to be classified is determined is determined.

Optionally, the step of determining the mask list according to the loss list and the true probability distribution includes:

Sort each value in the loss list to obtain a probability sequence, and select the target probability arranged in the previous preset position from the probability sequence;

determining the arrangement position of each target probability in the loss list, and updating the loss list according to the arrangement position;

The updated loss list and the true probability distribution are summed to generate a mask list.

Optionally, the step of determining the positive category loss and multiple negative category losses of the sample data according to the mask list and the loss list includes:

Perform a product operation on the mask list and the loss list before updating to generate a product result list;

According to the position of the true probability in the true probability distribution, the forward category loss is determined from the product result list;

According to the arrangement position, a plurality of the negative class losses are determined from the product result.

Optionally, the step of determining the loss list of the sample data on multiple preset categories according to the real probability distribution and the predicted probability distribution includes:

According to the real probability distribution and the predicted probability distribution, respectively determine the real probability value and the predicted probability value of the sample data on each preset category;

Based on a preset loss formula, calculate the real probability value and the predicted probability value of the sample data in each preset category, and obtain the loss value of the sample data in each preset category;

The loss list is obtained by arranging each of the loss values according to the arrangement order of each of the preset categories in the plurality of preset categories.

Optionally, the step of determining the predicted probability distribution of the sample data on multiple preset categories includes:

Obtain the output values corresponding to the sample data and a plurality of preset categories respectively, and map the output values respectively according to a preset function to obtain the predicted probability distribution of the sample data on the multiple preset categories .

Optionally, before the step of acquiring the real probability distribution of the sample data on a plurality of preset categories, the method further includes:

Obtain sample data and a sample label corresponding to the sample data, and encode the sample label based on a preset encoding method to generate a true probability distribution of the sample data on multiple preset categories.

Optionally, the step of determining, according to the positive class loss and a plurality of the negative class losses, to which the sample data belongs to a positive class and a plurality of negative classes in a plurality of preset classes includes:

train an initial model according to the positive class loss and a plurality of the negative class losses, and obtain gradients corresponding to the initial model;

When the gradient is less than a preset threshold, the training of the initial model is completed, and the forward category and the multiple preset categories are determined according to the mapping attributes of the sample data in multiple preset categories when the training of the initial model is completed. one of the negative categories.

Further, in order to achieve the above object, the present invention also provides a multi-category processing task training sample construction device, and the multi-category processing task training sample construction device includes:

an acquisition module, configured to acquire the real probability distribution of sample data on multiple preset categories, and determine the predicted probability distribution of the sample data on multiple preset categories;

a first determining module, configured to determine a loss list of the sample data in multiple preset categories according to the real probability distribution and the predicted probability distribution;

The second determination module is configured to determine a mask list according to the loss list and the true probability distribution, and determine the forward category loss and multiple values of the sample data according to the mask list and the loss list negative class loss;

a building module for determining, according to the positive category loss and a plurality of the negative category losses, a positive category and a plurality of negative categories to which the sample data belongs in a plurality of preset categories, and according to the A positive category and a plurality of negative categories are respectively constructed to construct positive and negative samples of a plurality of preset categories, so as to generate a multi-category classification model based on the positive and negative samples of the multiple preset categories for category classification.

Further, in order to achieve the above object, the present invention also provides a multi-class processing task training sample construction device, the multi-class processing task training sample construction device includes a memory, a processor, and a multi-class processing task training sample construction device. A multi-class processing task training sample construction program running on the processor, when the multi-class processing task training sample construction program is executed by the processor, implements the steps of the above-mentioned multi-class processing task training sample construction method.

Further, in order to achieve the above object, the present invention also provides a medium on which a multi-category processing task training sample construction program is stored, and the multi-category processing task training sample construction program is executed by the processor. The steps of the training sample construction method for the multi-class processing task.

Compared with the multi-class processing task training sample construction method, device and medium of the present invention, in the multi-class classification task in the prior art, the multi-class classification task is inaccurately performed due to the unbalanced sample data of each class. First, obtain the real probability distribution of sample data on multiple preset categories, and determine the predicted probability distribution of sample data on multiple preset categories; then, according to the real probability distribution and predicted probability distribution, determine the sample data in multiple preset categories Set the loss list on the category; then determine the mask list according to the loss list and the true probability distribution; and determine the positive category loss and multiple negative category losses of the sample data from the mask list and the loss list, and the positive category loss is determined by the positive category loss. Category loss and multiple negative category losses, determine the positive category and multiple negative categories that the sample data belongs to in multiple preset categories, and construct multiple presets according to the positive category and multiple negative categories. The positive and negative samples of the category are used to generate a multi-category classification model for category classification based on the positive and negative samples of multiple preset categories. Among them, the mask list reflects the positive probability that the sample data may belong to a certain category of multiple preset categories, and the negative probability that may not belong to some categories; it filters the loss list to obtain the positive category loss and multi-category loss. A negative category loss is used to determine the positive category and multiple negative categories that the sample data belongs to in multiple preset categories. Different sample data belong to different positive categories and negative categories, so that each preset category belongs to Each corresponds to its own positive samples and multiple negative samples, and positive and negative samples are constructed on multiple preset categories, so that each preset category contains positive samples and multiple negative samples to achieve preset categories. Balance between sample data. In this way, the defect of unbalanced sample data among various categories in the prior art multi-category classification task is overcome, and the accuracy of performing the multi-category classification task is improved.

Description of drawings

1 is a schematic structural diagram of a device hardware operating environment involved in an embodiment of a multi-category processing task training sample construction device according to the present invention;

2 is a schematic flowchart of a first embodiment of a method for constructing training samples for multi-category processing tasks according to the present invention;

FIG. 3 is a schematic diagram of functional modules of a preferred embodiment of an apparatus for constructing training samples for multi-class processing tasks according to the present invention.

The realization, functional characteristics and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The present invention provides a multi-category processing task training sample construction device. The multi-category processing task training sample construction device includes a wind control device and at least one consumer device communicatively connected to the wind control device. Referring to FIG. 1, FIG. 1 is a A schematic structural diagram of the device hardware operating environment involved in the multi-category processing task training sample construction device embodiment scheme of the present invention.

As shown in FIG. 1 , the multi-class processing task training sample construction device may include: a processor 1001 , such as a CPU, a communication bus 1002 , a user interface 1003 , a network interface 1004 , and a memory 1005 . Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (eg, a WI-FI interface). The memory 1005 may be high-speed RAM memory, or may be non-volatile memory, such as disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001 .

Those skilled in the art can understand that the hardware structure of the multi-class processing task training sample construction device shown in FIG. 1 does not constitute a limitation on the multi-class processing task training sample construction device, and may include more or less than the illustration. components, or a combination of certain components, or a different arrangement of components.

As shown in FIG. 1 , the memory 1005 as a medium may include an operating system, a network communication module, a user interface module, and a multi-class processing task training sample construction program. Among them, the operating system is a program that manages and controls the multi-class processing task training sample construction equipment and software resources, and supports the operation of the network communication module, the user interface module, the multi-class processing task training sample construction program and other programs or software; the network communication module Used to manage and control the network interface 1004; the user interface module is used to manage and control the user interface 1003.

In the hardware structure of the multi-category processing task training sample construction device shown in FIG. 1, the network interface 1004 is mainly used to connect to the background server and perform data communication with the background server; the user interface 1003 is mainly used to connect the client (client), and The client performs data communication; the processor 1001 can call the multi-class processing task training sample construction program stored in the memory 1005, and perform the following operations:

Obtain the real probability distribution of the sample data on the multiple preset categories, and determine the predicted probability distribution of the sample data on the multiple preset categories;

According to the real probability distribution and the predicted probability distribution, determine the loss list of the sample data in multiple preset categories;

Determine a mask list according to the loss list and the true probability distribution, and determine a positive category loss and a plurality of negative category losses of the sample data according to the mask list and the loss list;

According to the positive category loss and a plurality of the negative category losses, determine a positive category and a plurality of negative categories to which the sample data belongs in a plurality of preset categories, and according to the positive category and the multiple negative categories A multi-category classification model is generated based on the positive and negative samples of the multiple preset categories for category classification.

Further, after the step of constructing positive and negative samples of a plurality of preset categories respectively according to the positive category and a plurality of negative categories, the processor 1001 may call the multi-category processing task training samples stored in the memory 1005. Build the program, and do the following:

training a preset multi-category model based on a plurality of positive and negative item samples of the preset categories to generate a multi-category classification model;

When the data to be classified is received, the data to be classified is classified based on the multi-class classification model, and the class to which the data to be classified is determined is determined.

Further, the step of determining the mask list according to the loss list and the true probability distribution includes:

Sort each value in the loss list to obtain a probability sequence, and select the target probability arranged in the previous preset position from the probability sequence;

determining the arrangement position of each target probability in the loss list, and updating the loss list according to the arrangement position;

The updated loss list and the true probability distribution are summed to generate a mask list.

Further, the step of determining the positive category loss and multiple negative category losses of the sample data according to the mask list and the loss list includes:

Perform a product operation on the mask list and the loss list before updating to generate a product result list;

According to the position of the true probability in the true probability distribution, the forward category loss is determined from the product result list;

According to the arrangement position, a plurality of the negative class losses are determined from the product result.

Further, the step of determining the loss list of the sample data on multiple preset categories according to the real probability distribution and the predicted probability distribution includes:

According to the real probability distribution and the predicted probability distribution, respectively determine the real probability value and the predicted probability value of the sample data on each preset category;

Based on a preset loss formula, calculate the real probability value and the predicted probability value of the sample data in each preset category, and obtain the loss value of the sample data in each preset category;

The loss list is obtained by arranging each of the loss values according to the arrangement order of each of the preset categories in the plurality of preset categories.

Further, the step of determining the predicted probability distribution of the sample data on a plurality of preset categories includes:

Obtain the output values corresponding to the sample data and a plurality of preset categories respectively, and map the output values respectively according to a preset function to obtain the predicted probability distribution of the sample data on the multiple preset categories .

Further, before the step of obtaining the real probability distribution of sample data on multiple preset categories, the processor 1001 may call the multi-category processing task training sample construction program stored in the memory 1005, and perform the following operations:

Obtain sample data and a sample label corresponding to the sample data, and encode the sample label based on a preset encoding method to generate a true probability distribution of the sample data on multiple preset categories.

Further, the step of determining, according to the positive category loss and a plurality of the negative category losses, to which the sample data belongs in the multiple preset categories, includes:

train an initial model according to the positive class loss and a plurality of the negative class losses, and obtain gradients corresponding to the initial model;

When the gradient is less than a preset threshold, the training of the initial model is completed, and the forward category and the multiple preset categories are determined according to the mapping attributes of the sample data in multiple preset categories when the training of the initial model is completed. one of the negative categories.

The specific implementation of the multi-category processing task training sample construction device of the present invention is basically the same as the following embodiments of the multi-category processing task training sample construction method, and will not be repeated here.

The present invention also provides a method for constructing training samples for multi-category processing tasks.

Referring to FIG. 2 , FIG. 2 is a schematic flowchart of a first embodiment of a method for constructing training samples for multi-category processing tasks according to the present invention.

Embodiments of the present invention provide embodiments of a method for constructing training samples for multi-category processing tasks. It should be noted that although a logical sequence is shown in the flowchart, in some cases, the execution may be performed in a sequence different from that here. steps shown or described. Specifically, the multi-category processing task training sample construction method in this embodiment includes:

Step S10, obtaining the real probability distribution of the sample data on multiple preset categories, and determining the predicted probability distribution of the sample data on the multiple preset categories;

The multi-category processing task training sample construction method in this embodiment constructs positive and negative samples including positive samples and negative samples for each category of the multi-category classification task. In multi-class classification tasks, each sample belongs to only one of multiple classes. Each category involved in the multi-category classification task is preset as a preset category, and for each piece of sample data, a label representing the category to which it belongs is preset. Moreover, for each preset category, the attribution category of the sample data is represented by a real probability distribution, and the real probability distribution includes a plurality of probability values. For labels, the category of the sample data is a certain probability value, and for other categories, it is another probability value, and the multiple probability values form a true probability distribution as a whole; in this way, each probability value reflects the category of the sample data. For example, for C categories [1...i...c], if the sample data belongs to category i, and the category to which it belongs is represented by a probability value of 1, and the other categories are represented by a probability value of 0, then the true probability distribution formed is [0...1...0].

Understandably, the sample data is represented in the form of labels, and the real probability distribution exists in the form of numerical values, so it is necessary to convert the labels into numerical values to form the real probability distribution. Specifically, before the step of acquiring the real probability distribution of the sample data on the multiple preset categories, the method further includes:

Step a: Obtain sample data and a sample label corresponding to the sample data, and encode the sample label based on a preset encoding method to generate a true probability distribution of the sample data on multiple preset categories.

Further, the sample data and the sample labels carried by it are obtained, and for the obtained sample data, the real sample labels in a plurality of preset categories are encoded by a preset encoding method, and the sample data is obtained. true probability distribution over a preset class. Among them, the preset encoding method is preferably the one-hot method; the true probability distribution vector is obtained by the one-hot method, which may be called a one-hot vector. For the above C preset categories, the true probability distribution is a vector of length C, only the ith position of the true category i has a value of 1, and the values of other positions are 0.

Further, the real probability distribution of the generated sample data on the multiple preset categories is acquired, and the predicted probability distribution of the sample data on the multiple preset categories is determined. The prediction probability distribution represents the prediction of the sample data, and the probability of each preset category to which the sample data belongs is determined. Specifically, the step of determining the predicted probability distribution of the sample data on multiple preset categories includes:

Step b, obtaining output values corresponding to the sample data and a plurality of preset categories respectively, and mapping a plurality of the output values according to a preset function to obtain the sample data on the multiple preset categories. Predict probability distributions.

Furthermore, a model for prediction is preset, and the category to which the sample data belongs is predicted by the model, so as to obtain output values corresponding to a plurality of preset categories respectively. In addition, a preset function for performing numerical mapping on the output value is preset. Each of the generated output values is acquired, and each output value is mapped by a preset function to be between 0 and 1 to obtain each real number to form the predicted probability distribution of the sample data on multiple preset categories. The preset function is preferably a softmax function, that is, each output value is mapped to a real number between 0 and 1 through the softmax function to form a predicted probability distribution. Among them, the normalized sum of each mapped real number is 1 to ensure that the sum of the probabilities of the sample data on multiple categories is also 1, indicating that the sample data belongs to one of the multiple categories. For example, for the above C preset categories, the formed predicted probability distribution can be represented by [y1, y2...yi...yc].

Step S20, according to the real probability distribution and the predicted probability distribution, determine the loss list of the sample data in multiple preset categories;

Further, after obtaining the actual probability distribution and the predicted probability distribution of the sample data on each preset category, the loss list of the sample data on the multiple preset categories can be determined according to the actual probability distribution and the predicted probability distribution. The error of predicting sample data on multiple preset categories is represented by the loss list. Specifically, according to the real probability distribution and the predicted probability distribution, the step of determining the loss list of the sample data on multiple preset categories includes:

Step S21, according to the real probability distribution and the predicted probability distribution, respectively determine the real probability value and the predicted probability value of the sample data on each preset category;

Step S22, based on a preset loss formula, calculate the real probability value and the predicted probability value of the sample data in each preset category, and obtain the loss of the sample data in each preset category value;

Step S23 , arranging each of the loss values according to the arrangement order of each of the preset categories in the plurality of preset categories to obtain the loss list.

Understandably, the real probability distribution includes the real probability values of the sample data on each preset type, and the predicted probability distribution includes the predicted probability values of the sample data on each preset type. For each preset category, the real probability value of the sample data is determined from the real probability distribution, and the predicted probability value of the sample data is determined from the predicted probability distribution. Furthermore, the actual probability value and the predicted probability value of the sample data in each preset category are calculated through a preset preset loss formula, such as the cross-entropy formula, to obtain the loss value of the sample data in each preset category. For example, for the ith category in the C preset categories, it is determined that the value of the ith position of the sample data in the real probability distribution is y, and the value of the ith position in the predicted probability distribution is y', then determine the sample data. The real probability value of the data in the ith category is y, and the predicted probability value is y', and then the preset loss formula is called to calculate y and y', and the loss value of the sample data in the ith category is obtained. Among them, the preset preset loss formula includes a basic formula and a deformation formula to the basic formula, and the basic formula is shown in the following formula (1):

Among them, L represents the loss value, y represents the true probability value, and y' represents the predicted probability value.

The above formula (1) is deformed, and the deformation formulas are obtained as shown in the following formulas (2), (3) and (4):

CE(p _t )=-log(p _t ) (4);

Among them, CE(p _t ), pt _{, and CE(p t )} all represent the loss value, and p represents the predicted probability value.

Through the above deformation formula, the expression of the loss value of the i-th category in the preset category of the sample data is obtained as shown in the following formula (5):

CE(pi _,t )=-log(pi _,t ) (5);

Further, the loss values calculated for each preset category are arranged according to the arrangement order of each preset category in the multiple preset categories, and the obtained sequence is the sample data on the multiple preset categories. list of losses. Among them, the expression of the loss list is shown in the following formula (6):

SoftmaxCrossEntropyLossList=[-log(p _1,t ),...,-log(pi _,t ),...,-log(p _c,t )](6);

Step S30: Determine a mask list according to the loss list and the true probability distribution, and determine the positive category loss and multiple negative category losses of the sample data according to the mask list and the loss list ;

Understandably, because the loss list represents the size of the error, the larger the error, the less likely the sample data belongs to the preset category. Therefore, the negative samples of the preset category can be determined by the size of the possibility, and the sample data can be formed into negative samples of the preset category with low possibility. After being generated as a loss list, a mask list is generated by combining the loss list with the true probability distribution. The mask list contains values representing negative samples from the loss list and values representing positive samples from the true probability distribution. After the mask list is determined, the loss list is further processed through the mask list to generate the positive category loss and multiple negative category losses of the sample data. The real category to which the sample data belongs in the multi-category classification task is represented by the positive category loss, that is, the positive sample; and the multiple negative category losses are used to represent the category that the sample data cannot belong to in the multi-category classification task, that is, the negative sample.

Step S40, according to the positive category loss and a plurality of the negative category losses, determine a positive category and a plurality of negative categories to which the sample data belongs in a plurality of preset categories, and according to the positive category category and multiple negative categories, respectively construct positive and negative samples of multiple preset categories, so as to generate a multi-category classification model based on the positive and negative samples of multiple preset categories for category classification.

Further, positive and negative samples of multiple preset categories are constructed according to the preset category to which the loss of the positive category is represented, and the multiple negative categories that represent the preset category that cannot be assigned. According to the forward category loss, the category to which the sample data belongs in the plurality of preset categories is determined, and the category to which the sample data belongs is the forward category, and the sample data is then formed into a forward sample of the assigned forward category. At the same time, according to the losses of multiple negative categories, determine the category that the sample data cannot belong to among the multiple preset categories, and the impossible category is the negative category, and then the sample data is formed into an impossible negative category. Negative samples for the class. After processing a large amount of sample data, the positive category to which the large amount of sample data belongs can be determined to form positive samples of each positive category, and each impossible negative category to form the negative category of each negative category. oriented samples, so that each preset category contains positive samples and multiple negative samples, so that the positive and negative samples of multiple preset categories can be constructed according to the positive category and multiple negative categories, ensuring that each preset Balance of samples between classes. After that, the preset multi-category model is trained by constructing positive and negative samples of each preset category, and the multi-category classification model is obtained to perform the multi-category classification task for category classification. accurate classification.

Further, the step of determining, according to the positive class loss and a plurality of the negative class losses, to which the sample data belongs to a plurality of preset classes, includes:

Step c1, according to the positive category loss and a plurality of the negative category losses, train the initial model, and obtain the gradient corresponding to the initial model;

Step c2, when the gradient is less than a preset threshold, complete the training of the initial model, and determine the forward direction according to the mapping attributes of the sample data in multiple preset categories when the initial model completes the training category and a plurality of said negative categories.

Further, the construction of training samples for multi-category processing tasks in this embodiment is implemented in the dynamic process of model training. According to the determined positive class loss and a plurality of negative class losses, an initial model preset for performing a multi-class classification task is trained, and the gradient of the initial model after training is obtained. Thereafter, the gradient is compared with a preset preset threshold to determine whether the gradient is less than the preset threshold, and if it is less than the initial model training is completed. On the contrary, if it is determined by comparison that the gradient is not less than the preset threshold, the training continues, and the training of the initial model is not completed until it is less than the preset threshold. The training samples in the initial model after the training is completed makes the initial model show convergence characteristics, and the classification task has a good effect; therefore, each training sample is used as a mapping according to the category to which it belongs and the category that does not belong to it. attribute, and then determine the positive category and multiple negative categories according to the mapping attribute; take the category to which the sample data belongs when the initial model training is completed as the positive category, and the category to which the sample data does not belong as the negative category. Further, the positive samples and the negative samples are divided according to the positive category and the plurality of negative categories, and the positive and negative samples of each preset category are constructed, which is beneficial to the balance of the samples among the preset categories.

Further, after the step of respectively constructing positive and negative samples of a plurality of preset categories according to the positive category and a plurality of the negative categories, the method further includes:

Step d1, based on a plurality of positive and negative item samples of the preset categories, train a preset multi-category model to generate a multi-category classification model;

Step d1, when receiving the data to be classified, classify the data to be classified based on the multi-class classification model, and determine the class to which the data to be classified belongs.

Furthermore, a preset multi-category model for training is preset, so as to generate a multi-category classification model by training the preset multi-category model. Specifically, the constructed positive and negative samples on each preset category are transmitted to the preset multi-category model, and the preset multi-category model is trained. The preset multi-category model includes a training end condition. When the trained preset multi-category model reaches the training end condition, the training is ended, and a multi-category classification model is generated for category classification.

Further, when the data to be classified is received, indicating that there is a need to perform a multi-class classification task, the multi-class classification model is called to classify the data to be classified, determine the category to which the data to be classified belongs, and divide the data to be classified into the In the attribution category, the accurate classification of the data to be classified is realized.

Compared with the multi-class processing task training sample construction method of the present invention, in the multi-class classification task in the prior art, the sample data of each class is unbalanced, resulting in inaccurate execution of the multi-class classification task. The present invention first obtains the sample data. The real probability distribution on multiple preset categories, and determine the predicted probability distribution of sample data on multiple preset categories; and then according to the real probability distribution and predicted probability distribution, determine the sample data on multiple preset categories. Loss list; then determine the mask list according to the loss list and the true probability distribution; and determine the positive category loss and multiple negative category losses of the sample data from the mask list and the loss list, and the positive category loss and the multiple negative category losses are determined. A negative category loss is used to determine the positive category and multiple negative categories that the sample data belongs to in multiple preset categories, and construct the positive and negative categories of multiple preset categories according to the positive category and multiple negative categories. directional samples, so as to generate a multi-category classification model for category classification based on positive and negative samples of multiple preset categories. Among them, the mask list reflects the positive probability that the sample data may belong to a certain category of multiple preset categories, and the negative probability that may not belong to some categories; it filters the loss list to obtain the positive category loss and multi-category loss. A negative category loss is used to determine the positive category and multiple negative categories that the sample data belongs to in multiple preset categories, and different sample data belong to different positive categories and negative categories, so that each preset category belongs to Each corresponds to its own positive samples and multiple negative samples, and positive and negative samples are constructed on multiple preset categories, so that each preset category contains positive samples and multiple negative samples to achieve preset categories. Balance between sample data. In this way, the defect of unbalanced sample data among various categories in the prior art multi-category classification task is overcome, and the accuracy of performing the multi-category classification task is improved.

Further, based on the first embodiment of the method for constructing training samples for multi-category processing tasks of the present invention, a second embodiment of the method for constructing training samples for multi-category processing tasks of the present invention is proposed.

The difference between the second embodiment of the multi-category processing task training sample construction method and the multi-category processing task training sample construction method of the first embodiment is that the mask is determined according to the loss list and the true probability distribution. The steps to list include:

Step S31, sort each value in the loss list to obtain a probability sequence, and select the target probability arranged in the previous preset position from the probability sequence;

Step S32, determining the arrangement position of each of the target probabilities in the loss list, and updating the loss list according to the arrangement position;

Step S33, performing a sum operation on the updated loss list and the true probability distribution to generate a mask list.

This embodiment determines the mask list in combination with the loss list and the true probability distribution. Specifically, a sorted operation is performed on each value in the loss list, and through the sorted operation, each value in the loss list is sorted in descending order to obtain a probability sequence. After that, the topn operation is performed on the probability sequence. The topn operation is to find the top n data in the sequence, and n is a hyperparameter set in advance according to the requirements. As a preset bit, it is usually set to a value less than or equal to 5. Therefore, through the topn operation, the numerical value arranged in the previous preset position is selected from the probability sequence as the target probability. If n is set to 3, select the first three values as the target probability to form a sublist: [-log(p _j,t ),-log(p _k,t ),-log(p _l,t )], where j, k, l represent the position of the loss list from which the three values come from. The position of the loss list from which each target probability comes from is determined as the arrangement position, and the loss list is updated according to the arrangement position. The updating process is to set the value of the arrangement position to a certain value, such as a value of 1; and set the value of other positions other than the arrangement position to another value, such as a value of 0. For example, for the preset category C equal to 10, that is, 10 preset categories, the arrangement positions of j, k, and l are the 3rd, 4th, and 8th positions, respectively, and the loss value of these three positions in the loss list is the largest. ; then the values of the three positions are set to 1, and the values of other positions are set to 0, and the resulting update loss list is [0, 0, 1, 1, 0, 0, 0, 1, 0, 0].

It should be noted that the sorting and selection operations performed on the loss list in this embodiment, that is, the sorted operation and the topn operation, can be represented by the following formula (7), and the formula (7) is:

SoftmaxCrossEntropyLossList=topn(sorted(SoftmaxCrossEntropyLossList))(7);

Further, the updated loss list and the true probability distribution are added together to generate a mask list. In addition to the values of the updated loss list, the mask list also contains the true probability values in the true probability distribution. For example, for the loss list whose values are [0, 0, 1, 1, 0, 0, 0, 1, 0, 0] after the above update, if the sixth digit of the true probability distribution is the true probability, then the true probability The distribution [0, 0, 0, 0, 0, 1, 0, 0, 0, 0] and the updated loss list [0, 0, 1, 1, 0, 0, 0, 1, 0, 0] do Add and get a masked list of [0, 0, 1, 1, 0, 1, 0, 1, 0, 0].

Further, according to the mask list and the loss list, the step of determining the positive class loss and multiple negative class losses of the sample data includes:

Step S34, performing a product operation on the mask list and the loss list before updating to generate a product result list;

Step S35, according to the position of the true probability in the true probability distribution, determine the forward category loss from the product result list;

Step S36, according to the arrangement position, determine a plurality of the negative class losses from the product result.

Further, a product operation is performed on the mask list and the loss list before updating, and the result obtained by multiplying the two is used as the product result list. Among them, the product operation is to multiply the values in the same position in the list, for example, the first value in the mask list is multiplied by the first value in the loss list before the update. Since the value of each position in the mask list is a value of 1 or a value of 0, during multiplication, the invalid negative samples in the loss list are filtered by the multiplication of the value of 0, and the loss is saved by the multiplication of the value of 1. Positive samples in the list as well as valid samples. And, the product operation can be expressed by the following formula (8), and the formula (8) is:

FinalLossList=MaskedLossList*SoftmaxCrossEntropyLossList(8);

Among them, FinalLossList represents the product result list, and MaskedLossList represents the mask list.

Further, after the filtered product result list is generated, the forward category loss is determined from the product result list according to the position of the true probability in the true probability distribution. For example, in the case where the 6th digit in the above true probability distribution is the true probability, the value arranged in the 6th digit in the product result list is obtained as the forward class loss. At the same time, according to the arrangement positions of the target probabilities arranged in the previous preset positions, a plurality of negative class losses are determined from the product results. For example, for the situation where the above-mentioned target probabilities are arranged at the 3rd, 4th, and 8th positions, the values in the 3rd, 4th, and 8th positions in the product result list are obtained as the negative direction. Category loss.

In this embodiment, the loss list is updated by sorting and selecting the loss list, and a mask list is generated in combination with the updated loss list and the true probability; and the mask list is used to filter the loss list before the update to generate a product result A list to determine the positive class loss and multiple negative class losses. In this way, positive and negative samples are constructed through the obtained positive category loss and multiple negative category losses; through the processing and construction of multiple sample data, each preset category contains multiple positive and negative samples, which is beneficial to The balance of samples of each preset category.

The invention also provides a multi-class processing task training sample construction device.

Referring to FIG. 3 , FIG. 3 is a schematic diagram of functional modules of the first embodiment of the apparatus for constructing training samples for multi-category processing tasks according to the present invention. The multi-category processing task training sample construction device includes:

an acquisition module 10, configured to acquire the real probability distribution of the sample data on multiple preset categories, and determine the predicted probability distribution of the sample data on multiple preset categories;

a first determination module 20, configured to determine a loss list of the sample data in multiple preset categories according to the real probability distribution and the predicted probability distribution;

The second determination module 30 is configured to determine a mask list according to the loss list and the true probability distribution, and determine the forward class loss and multivariate loss of the sample data according to the mask list and the loss list negative class loss;

The building module 40 is configured to determine, according to the positive category loss and a plurality of the negative category losses, a positive category and a plurality of negative categories to which the sample data belongs in a plurality of preset categories, and according to the The positive and negative categories are described above, and positive and negative samples of a plurality of preset categories are respectively constructed, so as to generate a multi-category classification model based on the positive and negative samples of the multiple preset categories for category classification.

Further, the multi-category processing task training sample construction device includes:

a training module for training a preset multi-category model based on a plurality of positive and negative item samples of the preset categories to generate a multi-category classification model;

A classification module, configured to classify the data to be classified based on the multi-class classification model when receiving the data to be classified, and determine the class to which the data to be classified belongs.

Further, the second determining module 30 further includes:

a sorting unit, configured to sort each value in the loss list, obtain a probability sequence, and select the target probability arranged in the previous preset position from the probability sequence;

an update unit, configured to determine the arrangement position of each of the target probabilities in the loss list, and update the loss list according to the arrangement position;

A summation operation unit, configured to perform an addition operation on the updated loss list and the true probability distribution to generate a mask list.

Further, the second determining module 30 further includes:

a product operation unit, configured to perform a product operation on the mask list and the loss list before updating to generate a product result list;

a first determining unit, configured to determine the forward category loss from the product result list according to the position of the true probability in the true probability distribution;

The first determining unit is further configured to determine a plurality of the negative class losses from the product result according to the arrangement position.

Further, the first determining module 20 also includes:

a second determining unit, configured to respectively determine the real probability value and the predicted probability value of the sample data on each preset category according to the real probability distribution and the predicted probability distribution;

A calculation unit, configured to calculate the true probability value and the predicted probability value of the sample data in each preset category based on a preset loss formula, to obtain the sample data in each preset category loss value;

an arrangement unit, configured to arrange each of the loss values according to the arrangement order of each of the preset categories in the plurality of preset categories to obtain the loss list.

Further, the acquisition module 10 also includes:

an obtaining unit, configured to obtain output values corresponding to the sample data and a plurality of preset categories respectively, and map a plurality of the output values respectively according to a preset function to obtain the sample data in a plurality of preset categories The predicted probability distribution on .

Further, the multi-category processing task training sample construction device further includes:

an encoding module, configured to obtain sample data and sample labels corresponding to the sample data, and encode the sample labels based on a preset encoding method to generate the true probability distribution of the sample data on multiple preset categories .

Further, the building module 40 also includes:

a training unit, configured to train an initial model according to the positive class loss and a plurality of the negative class losses, and obtain gradients corresponding to the initial model;

A third determining unit, configured to complete the training of the initial model when the gradient is less than a preset threshold, and determine the mapping attributes of the sample data in multiple preset categories when the initial model completes the training the positive class and a plurality of the negative classes.

The specific implementations of the multi-class processing task training sample construction apparatus of the present invention are basically the same as the above-mentioned embodiments of the multi-class processing task training sample construction methods, and will not be repeated here.

In addition, an embodiment of the present invention also provides a medium.

A multi-class processing task training sample construction program is stored on the medium, and the multi-class processing task training sample construction program is executed by the processor to implement the steps of the multi-class processing task training sample construction method described above.

The medium of the present invention may be a computer medium, and its specific implementation is basically the same as the above-mentioned embodiments of the multi-category processing task training sample construction method, and will not be repeated here.

The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the inspiration of the present invention, without departing from the scope of protection of the purpose of the present invention and the claims, many forms can be made. Directly or indirectly applied in other related technical fields, these all belong to the protection of the present invention.

Claims (10)

1. A multi-category processing task training sample construction method, wherein the multi-category processing task training sample construction method comprises the following steps: Obtain the real probability distribution of the sample data on the multiple preset categories, and determine the predicted probability distribution of the sample data on the multiple preset categories; According to the real probability distribution and the predicted probability distribution, determine the loss list of the sample data in multiple preset categories; Determine a mask list according to the loss list and the true probability distribution, and determine a positive category loss and a plurality of negative category losses of the sample data according to the mask list and the loss list; According to the positive category loss and a plurality of the negative category losses, determine a positive category and a plurality of negative categories to which the sample data belongs in a plurality of preset categories, and according to the positive category and the multiple negative categories A multi-category classification model is generated based on the positive and negative samples of the multiple preset categories for category classification. 2 . The method for constructing training samples for multi-category processing tasks according to claim 1 , wherein, according to the positive category and a plurality of negative categories, respectively construct a plurality of preset categories of positive and negative samples. 3 . After the step, the method further includes: training a preset multi-category model based on a plurality of positive and negative item samples of the preset categories to generate a multi-category classification model; When the data to be classified is received, the data to be classified is classified based on the multi-class classification model, and the class to which the data to be classified is determined is determined. 3. The multi-class processing task training sample construction method according to claim 1, wherein the step of determining the mask list according to the loss list and the true probability distribution comprises: Sort each value in the loss list to obtain a probability sequence, and select the target probability arranged in the previous preset position from the probability sequence; determining the arrangement position of each target probability in the loss list, and updating the loss list according to the arrangement position; The updated loss list and the true probability distribution are summed to generate a mask list. 4. The method for constructing training samples for multi-class processing tasks according to claim 3, wherein the positive class loss and multiple negative class losses of the sample data are determined according to the mask list and the loss list. The steps towards class loss include: Perform a product operation on the mask list and the loss list before updating to generate a product result list; According to the position of the true probability in the true probability distribution, the forward category loss is determined from the product result list; According to the arrangement position, a plurality of the negative class losses are determined from the product result. 5 . The method for constructing training samples for multi-class processing tasks according to claim 1 , wherein, according to the real probability distribution and the predicted probability distribution, the determination of the sample data on multiple preset classes is performed. 6 . The steps for the loss list include: According to the real probability distribution and the predicted probability distribution, respectively determine the real probability value and the predicted probability value of the sample data on each preset category; Based on a preset loss formula, calculate the real probability value and the predicted probability value of the sample data in each preset category, and obtain the loss value of the sample data in each preset category; The loss list is obtained by arranging each of the loss values according to the arrangement order of each of the preset categories in the plurality of preset categories. 6. The multi-class processing task training sample construction method according to any one of claims 1-5, wherein the step of determining the prediction probability distribution of the sample data on a plurality of preset classes comprises: Obtain the output values corresponding to the sample data and a plurality of preset categories respectively, and map the output values respectively according to a preset function to obtain the predicted probability distribution of the sample data on the multiple preset categories . 7. The method for constructing training samples for multi-category processing tasks according to any one of claims 1 to 5, wherein before the step of acquiring the true probability distribution of sample data on multiple preset categories, the method Also includes: Obtain sample data and a sample label corresponding to the sample data, and encode the sample label based on a preset encoding method to generate a true probability distribution of the sample data on multiple preset categories. 8. The multi-class processing task training sample construction method according to any one of claims 1-5, wherein the sample is determined according to the positive class loss and a plurality of the negative class losses The steps of a positive category and a plurality of negative categories that the data belongs to in a plurality of preset categories include: train an initial model according to the positive class loss and a plurality of the negative class losses, and obtain gradients corresponding to the initial model; When the gradient is less than a preset threshold, the training of the initial model is completed, and the forward category and the multiple preset categories are determined according to the mapping attributes of the sample data in multiple preset categories when the training of the initial model is completed. one of the negative categories. 9. A multi-class processing task training sample construction device, characterized in that, the multi-class processing task training sample construction device comprises a memory, a processor, and a plurality of A class processing task training sample construction program, when the multi-class processing task training sample construction program is executed by the processor, implements the steps of the multi-class processing task training sample construction method according to any one of claims 1-8. 10. A medium, characterized in that a multi-category processing task training sample construction program is stored on the medium, and the multi-category processing task training sample construction program is executed by a processor to implement any one of claims 1-8. The steps of the multi-category processing task training sample construction method described in item.

Images