WO2022156589A1 - Method and device for determining live broadcast click rate - Google Patents

Abstract

The present disclosure relates to the technical field of computers, and provides a method and device for determining a live broadcast click rate. A specific embodiment of the method comprises: acquiring a plurality of pieces of historical user data and a plurality of pieces of historical live broadcast data; according to a sequence generation model, the plurality of pieces of historical user data and the generation time of the plurality of pieces of historical user data, determining a user behavior sequence corresponding to the plurality of pieces of historical user data; determining user attribute features according to the plurality of pieces of historical user data, and determining live broadcast features according to the plurality of pieces of historical live broadcast data; according to the user behavior sequence, the user attribute features and the live broadcast features, training a click rate prediction model; and according to the trained click rate prediction model, determining a click rate of a target user about target live broadcast data. The embodiment can improve the prediction accuracy of a live broadcast click rate, thereby improving the pushing accuracy, and reducing the possibility of phenomena such as insufficient stock or unsalable stock.

Description

A kind of method and device for determining click-through rate of live broadcast

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the priority of the Chinese invention patent application with application number 202110081300.7 filed on January 21, 2021, the entire contents of which are incorporated herein by reference.

technical field

The present disclosure relates to the field of computer technology, and in particular, to a method and device for determining the click-through rate of a live broadcast.

Background technique

As an emerging industry, live streaming has gradually become a major marketing method in the e-commerce industry.

In the prior art, operators mainly predict users’ clicks on the live broadcast based on their own experience or market experience based on live broadcast data (such as the time period of the live broadcast, the number of fans of the anchor, whether there are stars, whether there is a lottery and the number of prizes, etc.). Rate.

Due to the limited experience of operators and the ever-changing data corresponding to different live broadcasts, the accuracy of the live broadcast click-through rate subjectively predicted based on the operator's individual experience is low. Inaccuracy, insufficient stock or slow sales.

SUMMARY OF THE INVENTION

In view of this, the embodiments of the present disclosure provide a method and apparatus for determining the click-through rate of live broadcast, which can improve the prediction accuracy of the click-through rate of live broadcast, thereby improving the accuracy of push, and reducing the possibility of insufficient stock or slow sales.

In order to achieve the above purpose, according to an aspect of the embodiments of the present disclosure, a method for determining a live broadcast click-through rate is provided, comprising:

Obtain multiple historical user data and multiple historical live broadcast data;

According to the sequence generation model, the multiple historical user data, and the generation time of the multiple historical user data, determine the user behavior sequence corresponding to the multiple historical user data;

Determine user attribute characteristics according to the plurality of historical user data, and determine the live broadcast characteristics according to the plurality of historical live broadcast data;

According to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics, the click rate prediction model is trained;

According to the trained click-through rate prediction model, the click-through rate of the target user on the target live broadcast data is determined.

Optionally, determining the user behavior sequence corresponding to the multiple historical user data according to the sequence generation model, the multiple historical user data, and the generation time of the multiple historical user data, including:

determining user behavior characteristics according to the plurality of historical user data;

The user behavior feature and the generation time corresponding to the user behavior feature are used as the input of the sequence generation model, and the weight value corresponding to each of the user behavior features is determined according to the output of the sequence generation model;

The user behavior sequence is generated according to the user behavior feature and the weight value.

Optionally, determining the weight corresponding to each of the user behavior features according to the output of the sequence generation model, including:

The output of the sequence generation model is normalized to obtain a weight value corresponding to each of the user behavior features.

Optionally, the training of the click-through rate prediction model according to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics includes:

The user behavior sequence is input into the ARMA model, and the user dynamic feature is determined according to the output of the ARMA model;

The user dynamic feature, the user attribute feature and the live broadcast feature are used as the input of the click-through rate prediction model to train the click-through rate prediction model.

Optionally, after determining the click-through rate of the target user on the target live broadcast data, the method further includes:

According to the click rate, the live broadcast data is pushed to the target user.

Optionally, after determining the click-through rate of the target user on the target live broadcast data, the method further includes:

According to the click rate, the inventory corresponding to the target live broadcast data is determined, and inventory management is performed according to the inventory.

Optionally, the sequence generation model is a random Senli model;

and / or,

The click-through rate prediction model is an XGBOOST model.

According to yet another aspect of the embodiments of the present disclosure, there is provided a device for determining a click-through rate of a live broadcast, including:

The acquisition module is used to acquire multiple historical user data and multiple historical live broadcast data;

a sequence generation module, configured to determine the user behavior sequence corresponding to the multiple historical user data according to the sequence generation model, the multiple historical user data, and the generation time of the multiple historical user data;

a feature generation module, configured to determine user attribute features according to the plurality of historical user data, and determine live broadcast features according to the plurality of historical live broadcast data;

a model training module, used for training the click-through rate prediction model according to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics;

The data processing module is used to determine the click rate of the target user on the target live broadcast data according to the trained click rate prediction model.

According to another aspect of the embodiments of the present disclosure, there is provided an electronic device for determining the click-through rate of a live broadcast, including:

one or more processors;

storage means for storing one or more programs,

When the one or more programs are executed by the one or more processors, the one or more processors implement the method for determining the click-through rate of a live broadcast provided by the present disclosure.

According to another aspect of the embodiments of the present disclosure, a computer-readable medium is provided, on which a computer program is stored, and when the program is executed by a processor, implements the method for determining the click-through rate of a live broadcast provided by the present disclosure.

The above-mentioned embodiments have the following advantages or beneficial effects: because the time-based user data training model is used to determine the technical means of the live broadcast click-through rate, the technical problems of inaccurate subjective prediction push, insufficient stock or unsalable phenomena are overcome, and the result is achieved. The technical effect of improving the prediction accuracy of the click-through rate of the live broadcast, thereby improving the accuracy of the push, and reducing the possibility of insufficient stock or slow sales.

Further effects of the above non-conventional alternatives will be described below in conjunction with specific embodiments.

Description of drawings

The accompanying drawings are used for a better understanding of the present disclosure, and do not constitute an improper limitation of the present disclosure. in:

FIG. 1 is an exemplary system architecture diagram of a method for determining a live click rate or an apparatus for determining a live click rate that is suitable for application in an embodiment of the present disclosure;

2 is a schematic diagram of a main process of a method for determining a live broadcast click-through rate according to an embodiment of the present disclosure;

3 is a schematic diagram of a detailed flow of a method for determining a live broadcast click-through rate according to an embodiment of the present disclosure;

4 is a schematic diagram of main modules of a device for determining a live click rate according to an embodiment of the present disclosure;

FIG. 5 is a schematic structural diagram of a computer system suitable for implementing a terminal device or a server according to an embodiment of the present disclosure.

Detailed ways

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to facilitate understanding and should be considered as exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted from the following description for clarity and conciseness.

ARMA model: Autoregressive moving average model (ie Autoregressive moving average model), is an important method to study time series, including: autoregressive model (Autoregressive model, referred to as AR model), moving average model (Moving average model, referred to as MA model) and an autoregressive moving average model (ARMA).

Censoring: refers to the property that the autocorrelation function (ACF) or partial autocorrelation function (PACF) of the time series is 0 after a certain order.

Tailing: refers to the property that the autocorrelation function (ACF) or partial autocorrelation function (PACF) of the time series is not all zero after a certain order.

AIC, Akaike Information Criterion (Akaike Information Criterion), is a standard to measure the goodness of statistical model fitting. Usually, the smaller the AIC value, the better the model.

BIC, Bayesian Information Criterion (Bayesian Information Criterion), is a standard to measure the goodness of the fitting of statistical models. Usually, the smaller the BIC value, the better the model. AIC is to select a good model from the perspective of prediction for prediction; while BIC is to select the model that best fits the data from the perspective of fitting

FIG. 1 shows an exemplary system architecture diagram of a method for determining a click-through rate of live broadcast or an apparatus for determining a click-through rate of a live broadcast that is suitable for use in an embodiment of the present disclosure. As shown in FIG. The exemplary system architecture of the method or the apparatus for determining the live click rate includes:

As shown in FIG. 1 , the system architecture 100 may include terminal devices 101 , 102 , and 103 , a network 104 and a server 105 . The network 104 is a medium used to provide a communication link between the terminal devices 101 , 102 , 103 and the server 105 . The network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others.

The user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages and the like. Various communication client applications may be installed on the terminal devices 101 , 102 and 103 , such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, and the like.

The terminal devices 101, 102, 103 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

The server 105 may be a server that provides various services, for example, a background management server that provides support for shopping websites browsed by the terminal devices 101 , 102 , and 103 . The background management server may analyze and process the received user feature query request and other data, and feed back the processing results (eg, user features) to the terminal devices 101 , 102 , and 103 .

It should be noted that the method for determining the click-through rate of live broadcast provided by the embodiment of the present disclosure is generally performed by the server 105 , and accordingly, the device for determining the click-through rate of the live broadcast is generally set in the server 105 .

It should be understood that the numbers of terminal devices, networks and servers in FIG. 1 are merely illustrative. There can be any number of terminal devices, networks and servers according to implementation needs.

FIG. 2 is a schematic diagram of the main process of a method for determining a live broadcast click-through rate according to an embodiment of the present disclosure. As shown in FIG. 2 , the method for determining a live broadcast click-through rate of the present disclosure includes:

Step S201, acquiring multiple historical user data and multiple historical live broadcast data.

Exemplarily, a plurality of historical user data and a plurality of historical live broadcast data are obtained based on the historical data of the platform. The historical user data may include data such as the user's age, gender, purchasing ability, occupation, and preference, and may also include the user's browsing data. , comments, favorites, add-ons, ordering, sharing and other operations data, can also include the time when the user operation data is generated; historical live broadcast data can include live broadcast brand, lottery, time, interaction, anchor and commodity information.

Step S202, according to the sequence generation model, the multiple historical user data and the generation time of the multiple historical user data, determine the user behavior sequence corresponding to the multiple historical user data.

Exemplarily, according to a plurality of historical user data acquired in step S201, the user behavior characteristics are determined based on the user operation data therein, and the corresponding generation time of the user behavior characteristics is determined based on the time when the user operation data is generated, and the user behavior characteristics and user behavior characteristics are determined. The corresponding generation time of the feature is input to the sequence generation model, and the feature score of each user behavior feature is output. The feature scores of the user behavior features are normalized to obtain the weight value corresponding to each user behavior feature; based on the data of the user behavior feature and the weight value corresponding to the user behavior feature, the user behavior sequence is generated after weighted summation. The user behavior sequence contains 12 elements, that is, the 12-month user behavior score, which characterizes the user's behavior in the past 1 year.

Step S203: Determine user attribute characteristics according to the plurality of historical user data, and determine live broadcast characteristics according to the plurality of historical live broadcast data.

Exemplarily, according to the plurality of historical user data obtained in step S201, the user attribute feature is determined based on the user information data therein, and the live broadcast feature is determined based on the plurality of historical live broadcast data therein.

Step S204 , train a click-through rate prediction model according to the user behavior sequence, the user attribute feature, and the live broadcast feature.

Exemplarily, according to the user behavior sequence obtained in step S202, the user behavior sequence is input into the ARMA model, the ARMA model is trained, and the ARMA model parameters are output as user dynamic features. The user dynamic characteristics, the user attribute characteristics obtained in step S203, and the live broadcast characteristics are input into the click-through rate prediction model, the click-through rate prediction model is trained, and the trained click-through rate prediction model is output. Among them, the CTR prediction model is the XGBOOST model.

Step S205, according to the trained click-through rate prediction model, determine the click-through rate of the target user on the target live broadcast data.

Exemplarily, the target user data and a plurality of live broadcast data to be pushed are acquired, and the click rate of the target user on the target live broadcast data is determined according to the target user data and the trained click rate prediction model. The click-through rate may include data of regular clicks, browsing, favorites, add-ons, and ordering. According to the click rate, push live broadcast data for the target user; or, according to the click rate, determine the inventory corresponding to the target live broadcast data, and carry out inventory management according to the inventory amount to timely increase, allocate supply warehouses or support demand warehouses.

In the embodiment of the present disclosure, by acquiring a plurality of historical user data and a plurality of historical live broadcast data; The user behavior sequence corresponding to the historical user data; the user attribute characteristics are determined according to the multiple historical user data, and the live broadcast characteristics are determined according to the multiple historical live broadcast data; according to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics , train the click-through rate prediction model; according to the trained click-through rate prediction model, determine the click rate of the target user on the target live broadcast data and other steps, which can adapt to the periodic changes of user behavior, optimize the performance of the live broadcast data prediction model, and make full use of Live broadcast resources, accurately predict live broadcast click-through rate, accurately push live broadcasts to users and manage inventory reasonably.

3 is a schematic diagram of a detailed process of a method for determining a live broadcast click-through rate according to an embodiment of the present disclosure. As shown in FIG. 3 , the method for determining a live broadcast click-through rate of the present disclosure includes:

In step S301, a database of live broadcast is constructed.

Exemplarily, a live broadcast database is constructed based on the existing historical data of the platform, and historical user data and historical live broadcast data are obtained from the historical data of the platform. Historical user data can include multiple pieces of data. Take historical user data of an e-commerce platform as an example. For example, historical user data can include data such as the user's age, gender, purchasing ability, occupation, and preferences. Alternatively, historical user data can be It includes data of users' operations such as browsing live broadcasts, browsing products, adding purchases, placing orders, sharing, and commenting. Alternatively, historical user data may include operation time data of users corresponding to the operation data. The historical live broadcast data may include multiple pieces. Taking historical user data of an e-commerce platform as an example, for example, the historical live broadcast data may include data such as the live broadcast brand, lottery, time, interaction, anchor, and merchandise.

Further, the platform can obtain relevant data on a regular basis, and update and save the obtained relevant data to the database.

Step S302, constructing a user attribute feature of the live broadcast.

Exemplarily, according to the live database constructed in step S301, the information data of the historical user data is extracted therefrom, and based on the information data of the historical user data, the user attribute feature can be determined. Take the historical user data of an e-commerce platform as an example. For example, the historical user data may include data such as the user's age, gender, occupation, preference category, purchasing ability, geographic location, and consumption time. The attribute characteristics may include user age characteristics, user gender characteristics, user occupation characteristics, user preference category characteristics, user purchasing ability characteristics, user geographic location characteristics, user consumption time characteristics, and the like.

In step S303, a live broadcast feature of the live broadcast is constructed.

Exemplarily, according to the live broadcast database constructed in step S301, information data of the historical live broadcast data is extracted therefrom, and based on the information data of the historical live broadcast data, the live broadcast feature can be determined. Take the historical user data of an e-commerce platform as an example. For example, the historical live broadcast data may include data such as the brand, lottery, time, interaction, anchor and product of the live broadcast, and the live broadcast characteristics determined based on the historical live broadcast data may include the live broadcast brand characteristics. , live sweepstakes features, live broadcast time features, live broadcast anchor features, live broadcast commodity features, and live broadcast additional features, etc.

Further, the characteristics of live broadcast brands include the characteristics of the number of live broadcast brands, the characteristics of the number of fans of the live broadcast brand, etc.; the characteristics of live broadcast lottery draws include the characteristics of whether the live broadcast draws lottery, the characteristics of the number of live broadcast lottery draws, etc.; the characteristics of live broadcast time include whether the live broadcast time is a weekend, the live broadcast time period, etc. The characteristics of live broadcast anchors include the characteristics of the number of live broadcast anchors, the characteristics of the number of fans of the live broadcast anchor, the characteristics of the type of live broadcast anchors, the characteristics of the type of goods brought by the live broadcast anchor, etc. The characteristics of live broadcast products include the characteristics of the number of live broadcast products, the characteristics of the average price of live broadcast products, and the characteristics of the type of live broadcast products. etc.; the additional features of live broadcast include whether there are star features in the live broadcast room, whether the live broadcast has continuous microphone features, etc.

Further, the user attribute feature of the live broadcast obtained based on step S302 and the live broadcast feature of the live broadcast obtained in step S302 include discrete features and continuous features. Discrete features (for example, user's age feature, gender feature, geographical location feature, etc.) have a certain sparsity, which will lead to a sharp drop in model performance. Therefore, it is necessary to perform Embedding processing on discrete features and output continuous vectors, so that the features It has better expressive power in feature space. Among them, the Embedding embedding process can extract features from the original data, and perform dimension reduction processing through the principle of matrix multiplication. Continuous features are inherently continuous and therefore do not require processing.

Further, input the processed continuous live broadcast user attribute features and live broadcast features into the XGBOOST model for training, and output the feature score of each feature. The higher the score, the higher the importance; the selected feature score is above the predetermined value. The characteristics of the final live broadcast are used as the user attribute characteristics and live broadcast characteristics. As an ensemble learning model, XGBOOST can alleviate errors caused by feature sparseness and correlation, effectively remove redundant features, and improve feature quality. The feature crossover capability of the XGBOOST model itself can be used in python to use the feature importance function of the model for feature screening. Because the focus of the two indicators of the product browsing rate in the live broadcast room and the product order rate in the live broadcast room are different, the importance of the output features is also different for the model training of the two indicators.

Step S304, constructing the user behavior characteristics of the live broadcast.

Exemplarily, according to the database constructed in step S301, the operation data of the historical user data is extracted therefrom, and based on the operation data of the historical user data, the user behavior characteristic can be determined. Taking the historical user data of an e-commerce platform as an example, for example, the historical user data may include the user's data on operations such as browsing live broadcasts, browsing products, adding purchases, placing orders, and sharing, and the user behavior characteristics determined based on the historical user data may include User browsing live broadcast features, user browsing product features, user add-on purchase features, user ordering features, user sharing features, etc.

Step S305, constructing a user behavior sequence of the live broadcast.

Exemplarily, according to the live database constructed in step S301, the operation time data corresponding to the operation data of the historical user data is extracted therefrom, and based on the operation time data corresponding to the operation data of the historical user data, the behavioral characteristics of each user can be determined. generation time.

Exemplarily, the user behavior characteristics of the live broadcast constructed according to step S304 and the corresponding generation time input sequence generation model of each user behavior characteristic are trained, and the feature score of each user behavior characteristic is output after calculating the information gain, and the feature The scores are normalized to obtain the weight of each user's behavioral feature, and the user's behavioral sequence can be obtained by summing the data of the user's behavioral feature and the weight of each user's behavioral feature. Among them, the weight can represent the importance/degree of importance of different features, and the sequence refers to a vector composed of the values of multiple features within a predetermined time period. The sequence generation model may be a random forest model.

Exemplarily, the user behavior characteristics of the live broadcast constructed according to step S304 include user browsing live broadcast characteristics, user browsing commodity characteristics, user add-on purchase characteristics, user ordering characteristics, and user sharing characteristics; the history is extracted from the live broadcast database constructed in step S301. The operation time data corresponding to the operation data of the user data determines the generation time of each user behavior feature, that is, the generation time of the user’s live broadcast feature, the user’s product browsing feature, the user’s additional purchase feature, the user’s order feature, and the user’s sharing feature. Among them, the feature of user browsing live broadcast is represented by A, the feature of user browsing products is represented by B, the feature of user add-on purchase is represented by C, the feature of user placing an order is represented by D, and the feature of user sharing is represented by E.

Exemplarily, a live broadcast database is constructed based on the historical data of the platform within the past one year, and historical user data and historical live broadcast data are obtained from the live broadcast database. Based on the acquired historical user data and historical live broadcast data, construct user behavior characteristics A, B, C, D, and E for the past year, and generate user behavior characteristics A, B, C, D, E and corresponding user behavior characteristics. Time input random forest model for training, calculate the information gain and output the feature scores of A, B, C, D, E, normalize the feature scores of A, B, C, D, E to obtain A, B, The weights W _A , W _B , W _C , W _D , and W _E of C, D, and E combine the data of A, B, C, D, and E with the weights W _A , W of A, B, C, D, and E. _B , _WC , _WD , and _WE perform weighted summation to obtain the user behavior sequence.

Exemplarily, the data of A, B, C, D, and E may be the number of times users browse live broadcasts, the times users browse products, the times users add purchases, the times users place orders, and the times users share. The weighted sum of the data of A, B, C, D, and E and the weights of _A , _B , _C , _D , and _E is carried out to obtain the user behavior sequence, including: The weighted sum of the data of A, B, C, D, and E in January and the weights of A, B, C, D, and E, W _A , W _B , W _C , W _D , and W _E , get the user in January Behavior scores; weighting the data of A, B, C, D, _E from February to December with the weights of _A , _B , _C , _D , E, WA, WB, WC, WD, WE, respectively Sum up to get the behavior scores of users from February to December; combine the user behavior scores of 12 months to obtain the user behavior sequence. The user behavior sequence contains 12 elements, that is, the 12-month user behavior score, which characterizes the user's behavior in the past 1 year.

Further, for example, input the data of A, B, C, D, and E into the random forest model, the output weights W _A , W _B , W _C , W _D , W _E are [0.3, 0.2, 0.2, 0.1, 0.2]; Compare the data of A, B, C, D, and E in January (users browsed the live broadcast 5 times, browsed products 10 times, added 3 times, placed an order 2 times, and shared 1 time) with A, B, C , D and E weights W _A , W _B , W _C , W _D , and W _E are weighted and summed to obtain the user’s behavior score in January s1=0.3*5+0.2*10+0.2*3+0.1*2+ 0.2*1=4.5; respectively compare the data of A, B, C, D, and E from February to December with the weights of A, B, C, D, and E W _A , W _B , W _C , W _D , W _E Perform the weighted summation to get the user behavior scores s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12 from February to December; combine the user behavior scores for 12 months to get User behavior sequence [s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12].

Further, due to the difference in the focus of the two indicators, the product viewing rate in the live broadcast room and the product order rate in the live broadcast room, the product viewing rate in the live broadcast room focuses on the pageviews, and the live broadcast room product order rate focuses more on the order volume. Therefore, the model training for the two metrics, the feature importance of the output is also different. For example, for the product browsing rate indicator in the live broadcast room, the data of A, B, C, D, and E are input into the random forest model, and the output weights W _A , W _B , W _C , W _D , W _E are [0.3, 0.2 , 0.2, 0.1, 0.2]; and for the product order rate indicator in the live broadcast room, the data of A, B, C, D, and E are input into the random forest model, and the output weights W _A , W _B , W _C , W _D , W _E is [0.1, 0.2, 0.2, 0.3, 0.2]. It can be seen that, for the product viewing rate index in the live broadcast room, the weights W _A and W _B of A and B are greater than the weight W _D of D, indicating the importance of user browsing live broadcast characteristics and user browsing product characteristics in the product viewing rate index in the live broadcast room. Higher than the user's ordering characteristics; for the live broadcast room product viewing rate and live broadcast room product ordering rate indicators, the weight of D W _D accounts for a higher proportion of the live broadcast room product ordering rate, indicating that the user's ordering characteristics are under the live broadcast room products. The importance is higher in the single rate indicator.

Step S306, constructing a user dynamic feature of the live broadcast.

Exemplarily, according to the live user behavior sequence constructed in step S305, the user behavior sequence is input into the ARMA model, and the stationarity of the user behavior sequence is detected by the ADF test. Based on the results of the ADF test, it is judged whether the user behavior sequence is stationary, and if not, the difference processing (difference operation) is performed until the user behavior sequence is stationary. After confirming that the user behavior sequence is stable, calculate the autocorrelation coefficient a and partial autocorrelation coefficient b of the user behavior sequence, and identify the ARMA model according to the autocorrelation coefficient a (ACF) and the partial autocorrelation coefficient b (PACF). Among them, identifying the ARMA model includes: if the autocorrelation coefficient a is tailing and the partial autocorrelation coefficient b is p-order truncation, then the ARMA model is an ARp model; if the autocorrelation coefficient a is q-order truncation, the partial autocorrelation coefficient b If the autocorrelation coefficient a is tailing and the partial autocorrelation coefficient b is also tailing, the ARMA model is the ARMAp,q model. Based on the determined ARMA model, the order p and q are determined in combination with AIC and BIC criteria. The order p and q represent the autocorrelation characteristics of the sequence itself, especially the periodic behavior. After the order p and q are determined, the model parameters of the ARMA model can be obtained, and the model parameters of the ARMA model are combined into a feature vector, which is the user dynamic feature.

Further, the user's periodic behavior, such as weekly and monthly inertial behavior, makes the user's behavior have periodic laws. User dynamic features are sequence features, which themselves are continuous.

Further, deep learning models such as RNN, LSTM, and temporal convolutional network can be used to construct user dynamic features.

Exemplarily, when the user behavior sequence is used as the input of the click-through rate prediction model of the live broadcast, there are problems such as too large dimension, which may cause waste of resources, and at the same time, the dimension and sequence length are not uniform. Therefore, the ARMA model is used to abstract the user behavior sequence, and its model parameters are used to construct the user dynamic characteristics of the live broadcast.

Step S307, the click rate prediction model is trained.

Exemplarily, according to the user dynamic characteristics of the live broadcast constructed in step S306, the user attribute characteristics of the live broadcast constructed in step S302, and the live broadcast characteristics constructed in step S303, a characteristic sample of the live broadcast click rate is constructed. The feature samples of the live CTR are divided into training set and test set, and 80% of the feature samples are selected as the training set for training the model to obtain the trained CTR prediction model; the remaining 20% of the feature samples are used as the test set. Used to test the CTR prediction model after live training.

Exemplarily, the training set feature samples of the live click rate are input into the XGBOOST model, the weak classifiers are trained in an iterative loop, the multiple weak classifiers are iteratively integrated into a combined classifier, and the trained click rate prediction model is obtained.

Further, the click-through rate prediction model is used to predict the click-through rate of the live broadcast, wherein the click-through rate may include data such as regular clicks, browsing, favorites, add-ons, ordering, and sharing.

Further, the XGBOOST model (eXtreme Gradient Boosting) is a boosting tree model, iterative training is performed according to the input training set feature samples, the weak classifiers of each iteration are learned step by step, and the coefficients of the samples in the training set are updated according to the coefficients of the weak classifiers. Weight; the residuals between the results of fitting previous weak classifiers and the training set samples, and iteratively integrates multiple weak classifiers into a strong classifier to obtain a prediction model. As an ensemble learning model, XGBOOST has good learning performance. At the same time, as an optimized distributed gradient boosting model, XGBOOST model can consume less computing resources in a short time and obtain a prediction model with excellent performance. It combines the advantages of sexual awareness and cross-validation.

Further, models such as LR, random forest, GBDT, BP neural network and other algorithms can be used for model training.

Step S308, verifying the click-through rate prediction model.

Exemplarily, input the test set feature samples of the live broadcast click rate into the combined classifier obtained by training in step S307, that is, the trained click rate prediction model, output the user's live broadcast click rate, and calculate the click rate prediction model according to the test result. Error, judge whether the model error is higher than the error standard value, and modify the click-through rate prediction model according to the model error, so that the click-through rate prediction model meets the requirements.

Further, through the association relationship based on sequence features, the prediction accuracy of the prediction model is greatly improved, and the click-through rate of the live broadcast can be accurately predicted.

Step S309, the model is used.

Exemplarily, based on the click-through rate prediction model obtained in step S308, according to the obtained target user data and a plurality of live broadcast data to be pushed, the obtained target user data and a plurality of live broadcast data to be pushed are input into the click-through rate prediction model, and output. The click-through rate of the target live stream data. According to the click rate, the live broadcast data can be pushed to the target user; or, according to the click rate, the inventory corresponding to the target live broadcast data can be determined, and inventory management can be carried out according to the inventory amount, for example, direct allocation of inventory from the origin, allocation from the supply warehouse Inventory to demand warehouse, etc.

Further, the click-through rate may include data of regular clicks, browsing, favorites, add-ons, and ordering. If the click-through rate of an order for a certain product is high, the warehouse will increase the inventory of the product; if a user has a high click-through rate to browse, favorite or add a certain product, it will recommend the live broadcast of the product or similar products to the user. 's live broadcast.

In the embodiments of the present disclosure, by constructing a live broadcast database; constructing a live broadcast user attribute feature; constructing a live broadcast feature; constructing a live broadcast user behavior feature; constructing a live broadcast user behavior sequence; constructing a live broadcast user dynamic feature; click-through rate prediction Model training; verification of the click-through rate prediction model; model use and other steps can adapt to the cyclical changes in user behavior, optimize the performance of the live broadcast data prediction model, make full use of live broadcast resources, accurately predict the live broadcast click rate, and can accurately push the live broadcast to users and reasonably Manage inventory.

FIG. 4 is a schematic diagram of the main modules of the apparatus for determining the click-through rate of live broadcast according to an embodiment of the present disclosure. As shown in FIG. 4 , the apparatus 400 for determining the click-through rate of live broadcast of the present disclosure includes:

The obtaining module 401 is configured to obtain multiple historical user data and multiple historical live broadcast data.

Exemplarily, the obtaining module 401 obtains a plurality of historical user data and a plurality of historical live broadcast data based on the historical data of the platform. It can include data of users' browsing, comments, favorites, add-ons, ordering, sharing, etc., as well as the time when the user's operation data is generated; historical live broadcast data can include live broadcast brand, lottery, time, interaction, anchor and commodity and other information data.

The sequence generation module 402 is configured to determine the user behavior sequence corresponding to the multiple historical user data according to the sequence generation model, the multiple historical user data, and the generation time of the multiple historical user data.

Exemplarily, according to a plurality of historical user data acquired by the acquisition module 401, the user behavior characteristics are determined based on the user operation data therein, and the generation time corresponding to the user behavior characteristics is determined based on the time when the user operation data is generated, and the sequence is generated. The module 402 generates a model by inputting the user behavior feature and the generation time sequence corresponding to the user behavior feature, and outputs a feature score of each user behavior feature. The sequence generation module 402 normalizes the feature scores of the user behavioral features to obtain a weight value corresponding to each user behavioral feature; based on the data of the user behavioral feature and the weighted value corresponding to the user behavioral feature, the weighted sum is generated. User behavior sequence. The user behavior sequence contains 12 elements, that is, the 12-month user behavior score, which characterizes the user's behavior in the past 1 year.

The feature generation module 403 is configured to determine user attribute features according to the multiple historical user data, and determine the live broadcast feature according to the multiple historical live broadcast data.

Exemplarily, according to the plurality of historical user data obtained by the obtaining module 401, the feature generating module 403 determines the user attribute feature based on the user information data therein, and determines the live broadcast feature based on the plurality of historical live broadcast data therein.

The model training module 404 is configured to train the click rate prediction model according to the user behavior sequence, the user attribute feature and the live broadcast feature.

Exemplarily, according to the user behavior sequence obtained by the sequence generation module 402, the model training module 404 inputs the user behavior sequence into the ARMA model, trains the ARMA model, and outputs the ARMA model parameters as user dynamic features. The model training module 404 inputs the user dynamic features, the user attribute features and the live broadcast features obtained by the feature generation module 403 into the click-through rate prediction model, trains the click-through rate prediction model, and outputs the trained click-through rate prediction model. Among them, the CTR prediction model is the XGBOOST model.

The data processing module 405 is configured to determine the click rate of the target user on the target live broadcast data according to the trained click rate prediction model.

Exemplarily, the target user data and a plurality of live broadcast data to be pushed are obtained, and the data processing module 405 determines the click rate of the target user on the target live broadcast data according to the target user data and the trained click rate prediction model. The click-through rate may include data of regular clicks, browsing, favorites, add-ons, and ordering. According to the click rate, push live broadcast data for the target user; or, according to the click rate, determine the inventory corresponding to the target live broadcast data, and carry out inventory management according to the inventory amount to timely increase, allocate supply warehouses or support demand warehouses.

In the embodiments of the present disclosure, modules such as an acquisition module, a sequence generation module, a feature generation module, a model training module, and a data processing module can adapt to periodic changes in user behavior, optimize the performance of the live broadcast data prediction model, and make full use of live broadcast resources. , which can accurately predict the click-through rate of live broadcasts, accurately push live broadcasts to users and manage inventory reasonably.

FIG. 5 is a schematic structural diagram of a computer system suitable for implementing the terminal device according to the embodiment of the present disclosure. As shown in FIG. 5 , the computer system 500 of the terminal device according to the embodiment of the present disclosure includes:

A central processing unit (CPU) 501 can execute various appropriate actions and processes according to a program stored in a read only memory (ROM) 502 or a program loaded into a random access memory (RAM) 503 from a storage section 508 . In the RAM 503, various programs and data necessary for the operation of the system 500 are also stored. The CPU 501 , the ROM 502 , and the RAM 503 are connected to each other through a bus 504 . An input/output (I/O) interface 505 is also connected to bus 504 .

The following components are connected to the I/O interface 505: an input section 506 including a keyboard, a mouse, etc.; an output section 507 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc. ; and a communication section 509 including a network interface card such as a LAN card, a modem, and the like. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I/O interface 505 as needed. A removable medium 511, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 510 as needed so that a computer program read therefrom is installed into the storage section 508 as needed.

In particular, the processes described above with reference to the flowcharts may be implemented as computer software programs according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart. In such an embodiment, the computer program may be downloaded and installed from the network via the communication portion 509 and/or installed from the removable medium 511 . When the computer program is executed by the central processing unit (CPU) 501, the above-described functions defined in the system of the present disclosure are executed.

It should be noted that the computer-readable medium shown in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or a combination of any of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), fiber optics, portable compact disk read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. In this disclosure, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present disclosure, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device . Program code embodied on a computer readable medium may be transmitted using any suitable medium including, but not limited to, wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logical functions for implementing the specified functions executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It is also noted that each block of the block diagrams or flowchart illustrations, and combinations of blocks in the block diagrams or flowchart illustrations, can be implemented in special purpose hardware-based systems that perform the specified functions or operations, or can be implemented using A combination of dedicated hardware and computer instructions is implemented.

The modules involved in the embodiments of the present disclosure may be implemented in software or hardware. The described modules can also be set in the processor, for example, it can be described as: a processor includes an acquisition module, a sequence generation module, a feature generation module, a model training module, and a data processing module. Wherein, the names of these modules do not constitute a limitation on the module itself under certain circumstances. For example, the acquisition module can also be described as "a module for acquiring live broadcast data from a live broadcast platform".

As another aspect, the present disclosure also provides a computer-readable medium. The computer-readable medium may be included in the device described in the above-mentioned embodiments, or it may exist alone without being assembled into the device. The above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by a device, the device includes: acquiring a plurality of historical user data and a plurality of historical live broadcast data; generating a model according to the sequence, The plurality of historical user data and the generation time of the plurality of historical user data determine the user behavior sequence corresponding to the plurality of historical user data; determine user attribute characteristics according to the plurality of historical user data, and A plurality of historical live broadcast data determine the live broadcast characteristics; according to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics, the click-through rate prediction model is trained; according to the trained click-through rate prediction model, the target user's information about the target live broadcast data is determined. CTR.

According to the technical solutions of the embodiments of the present disclosure, the prediction accuracy of the click-through rate of the live broadcast can be improved, so that the push accuracy can be improved, and the possibility of insufficient stock or slow sales can be reduced.

The above-mentioned specific embodiments do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present disclosure should be included within the protection scope of the present disclosure.

Claims (10)

A method for determining the click-through rate of a live broadcast, including: Obtain multiple historical user data and multiple historical live broadcast data; According to the sequence generation model, the multiple historical user data, and the generation time of the multiple historical user data, determine the user behavior sequence corresponding to the multiple historical user data; Determine user attribute characteristics according to the plurality of historical user data, and determine the live broadcast characteristics according to the plurality of historical live broadcast data; According to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics, the click rate prediction model is trained; According to the trained click-through rate prediction model, the click-through rate of the target user on the target live broadcast data is determined. The method according to claim 1, wherein the user behavior sequence corresponding to the plurality of historical user data is determined according to a sequence generation model, the plurality of historical user data, and the generation time of the plurality of historical user data ,include: determining user behavior characteristics according to the plurality of historical user data; The user behavior feature and the generation time corresponding to the user behavior feature are used as the input of the sequence generation model, and the weight value corresponding to each of the user behavior features is determined according to the output of the sequence generation model; The user behavior sequence is generated according to the user behavior feature and the weight value. The method according to claim 2, wherein determining the weight corresponding to each of the user behavior characteristics according to the output of the sequence generation model, comprising: The output of the sequence generation model is normalized to obtain a weight value corresponding to each of the user behavior features. The method according to claim 1, wherein the training of the click-through rate prediction model according to the user behavior sequence, the user attribute feature and the live broadcast feature includes: The user behavior sequence is input into the ARMA model, and the user dynamic feature is determined according to the output of the ARMA model; The user dynamic feature, the user attribute feature, and the live broadcast feature are used as the input of the click-through rate prediction model, so as to train the click-through rate prediction model. The method according to claim 1, wherein after determining the click-through rate of the target user on the target live broadcast data, the method further comprises: According to the click rate, the live broadcast data is pushed to the target user. The method according to claim 1, wherein after determining the click-through rate of the target user on the target live broadcast data, the method further comprises: According to the click rate, the inventory corresponding to the target live broadcast data is determined, and inventory management is performed according to the inventory. The method of claim 1, wherein, The sequence generation model is a random forest model; and / or, The click-through rate prediction model is an XGBOOST model. A device for determining the click-through rate of live broadcast, comprising: The acquisition module is used to acquire multiple historical user data and multiple historical live broadcast data; a sequence generation module, configured to determine the user behavior sequence corresponding to the multiple historical user data according to the sequence generation model, the multiple historical user data, and the generation time of the multiple historical user data; a feature generation module, configured to determine user attribute features according to the plurality of historical user data, and determine live broadcast features according to the plurality of historical live broadcast data; a model training module, used for training the click-through rate prediction model according to the user behavior sequence, the user attribute characteristics and the live broadcast characteristics; The data processing module is used to determine the click rate of the target user on the target live broadcast data according to the trained click rate prediction model. An electronic device for determining the click-through rate of a live broadcast, comprising: one or more processors; storage means for storing one or more programs, The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 1-7. A computer-readable medium having a computer program stored thereon, the program implementing the method of any one of claims 1-7 when executed by a processor.

Images