Model training method, product click-through rate prediction method, system, equipment and medium

Abstract

The present invention discloses a model training method, a product click-through rate prediction method, a system, a device, and a medium. The model training method obtains data to be trained; selects prior feature data and non-prior feature data from feature data of a training user, feature data of a training product, and feature data of a training user's request; and inputs the prior feature data, non-prior feature data, and click interaction data into a first model for training to obtain a product prediction model. By adding a second neural network layer that strengthens the prior features, the present invention effectively helps the model better learn the prior features, solves the problem that the learning effect of strongly correlated prior features in the product prediction model is not obvious, enhances the prediction effect of the model, and effectively improves the prediction accuracy of the probability value of a product being clicked by a user.

Description

Technical Field

[0001] The present invention relates to a deep model ranking method, and in particular to a model training method, a product click-through rate prediction method, system, device and medium. Background Art

[0002] Ranking algorithms are widely used in search, recommendation, and advertising scenarios, enhancing user experience and increasing supplier revenue. Sorting algorithms are used to rank products in search, recommendation, and advertising scenarios, and the results are presented directly to users. This ranking determines both the user experience and the supplier's revenue.

[0003] The candidate products to be recalled are screened. In the process of sorting the screened target products, it is necessary to combine a large number of user characteristics and product characteristics to predict the probability value of the user's interactive behavior (click / exposure) on each target product, and sort the target products with which the user may have interactive behavior based on the probability value.

[0004] Currently, the predictive effectiveness of deep learning models primarily depends on the application scenario and the capabilities of the development team, and remains at the stage of mining feature interactions and user behavior sequences. This results in low prediction accuracy due to the model's poor learning of features that are strongly correlated with product clicks and impressions. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the defect in the prior art of low prediction accuracy caused by the poor learning effect of the model on features strongly correlated with product clicks / exposures, and to provide a model training method, a product click-through rate prediction method, system, equipment and medium.

[0006] The present invention solves the above technical problems through the following technical solutions:

[0007] In a first aspect, the present invention provides a model training method, comprising:

[0008] Acquire data to be trained; the data to be trained includes characteristic data of the training user, characteristic data of the training product, characteristic data of the training user's request, and click interaction data of the training user on the training product;

[0009] Filtering out priori feature data and non-priori feature data from the feature data of the training user, the feature data of the training product, and the feature data of the training user's request;

[0010] The prior feature data, the non-prior feature data and the click interaction data are input into a first model for training to obtain a product prediction model; the first model includes a first neural network layer and a second neural network layer that strengthens the prior features, and the product prediction model is used to calculate the probability value of each product being clicked by the user in future orders in a future time period.

[0011] Preferably, the second neural network layer for strengthening the prior features includes a first fully connected network unit, a second fully connected network unit, a Softmax activation function processing network unit and a processing network unit; the Softmax activation function processing network unit is located between the first fully connected network unit and the processing network unit;

[0012] The input of the Softmax activation function processing network unit includes the output of the first fully connected network unit, and the output of the Softmax activation function processing network unit includes the enhancement coefficient of each feature expression of the output of the second fully connected network unit;

[0013] The input of the processing network unit includes the output of the Softmax activation function processing network unit and the output of the second fully connected network unit, and the dimension of the feature expression of the output of the Softmax activation function processing network unit and the feature expression of the output of the second fully connected network unit are the same.

[0014] Preferably, the step of inputting the priori feature data, the non-priori feature data and the click interaction data into a first model for training to obtain a product prediction model includes:

[0015] After inputting the non-prior feature data into the first neural network layer and the prior feature data into the second neural network layer respectively, the first model is trained based on the click interaction data.

[0016] Preferably, the non-a priori feature data includes discrete feature data and continuous feature data, and the a priori feature data includes a priori discrete feature data and a priori continuous feature data;

[0017] After inputting the non-a priori feature data into the first neural network layer and the a priori feature data into the second neural network layer, the step of training the first model based on the click interaction data includes:

[0018] Preprocessing the discrete feature data and the prior discrete feature data using Embedding technology;

[0019] Processing the processed discrete feature data and the processed continuous feature data through at least two layers including a connection layer and a fully connected layer, respectively, to obtain a first combined feature expression;

[0020] Processing the processed prior discrete feature data and the processed prior continuous feature data respectively through at least two layers including a connection layer and a processing layer for strengthening prior features to obtain a second combined feature expression;

[0021] The first combined feature expression and the second combined feature expression are concatenated and sequentially processed through at least three layers including a connection layer, a fully connected layer, and an activation processing layer to obtain a predicted click probability of the training user for the training product;

[0022] The binary cross entropy loss function is fitted according to the predicted click probability and the click interaction data of the training users on the training product to generate the product prediction model.

[0023] Preferably, the discrete feature data includes at least one of article type, article ID, user gender, user occupation and weather conditions; the continuous feature data includes at least one of article popularity, user age and the number of articles browsed by the user per day; the a priori discrete feature data includes at least one of article payment status and whether the user recharges; the a priori continuous feature data includes at least one of the number of article comments, the number of article clicks and the number of times users comment on articles.

[0024] In a second aspect, the present invention further provides a model training system, comprising:

[0025] The first acquisition module is used to acquire data to be trained; the data to be trained includes characteristic data of the training user, characteristic data of the training product, characteristic data of the training user's request, and click interaction data of the training user on the training product;

[0026] A screening module, configured to screen out priori feature data and non-priori feature data from the feature data of the training user, the feature data of the training product, and the feature data requested by the training user;

[0027] A training module is used to input the prior feature data, the non-prior feature data and the click interaction data into a first model for training to obtain a product prediction model; the first model includes a first neural network layer and a second neural network layer that strengthens the prior features, and the product prediction model is used to calculate the probability value of each product being clicked by the user in future orders in a future time period.

[0028] Preferably, the second neural network layer for strengthening the prior features includes a first fully connected network unit, a second fully connected network unit, a Softmax activation function processing network unit and a processing network unit; the Softmax activation function processing network unit is located between the first fully connected network unit and the processing network unit;

[0029] The input of the Softmax activation function processing network unit includes the output of the first fully connected network unit, and the output of the Softmax activation function processing network unit includes the enhancement coefficient of each feature expression of the output of the second fully connected network unit;

[0030] The input of the processing network unit includes the output of the Softmax activation function processing network unit and the output of the second fully connected network unit, and the dimension of the feature expression of the output of the Softmax activation function processing network unit and the feature expression of the output of the second fully connected network unit are the same.

[0031] Preferably, the training module includes:

[0032] A training unit is used to input the non-prior feature data into the first neural network layer and the prior feature data into the second neural network layer, and then train the first model based on the click interaction data.

[0033] Preferably, the non-a priori feature data includes discrete feature data and continuous feature data, and the a priori feature data includes a priori discrete feature data and a priori continuous feature data;

[0034] The training unit is specifically used for:

[0035] Preprocessing the discrete feature data and the prior discrete feature data using Embedding technology;

[0036] Processing the processed discrete feature data and the processed continuous feature data through at least two layers including a connection layer and a fully connected layer, respectively, to obtain a first combined feature expression;

[0037] Processing the processed prior discrete feature data and the processed prior continuous feature data respectively through at least two layers including a connection layer and a processing layer for strengthening prior features to obtain a second combined feature expression;

[0038] The first combined feature expression and the second combined feature expression are concatenated and sequentially processed through at least three layers including a connection layer, a fully connected layer, and an activation processing layer to obtain a predicted click probability of the training user for the training product;

[0039] The binary cross entropy loss function is fitted according to the predicted click probability and the click interaction data of the training users on the training product to generate the product prediction model.

[0040] Preferably, the discrete feature data includes at least one of article type, article ID, user gender, user occupation and weather conditions; the continuous feature data includes at least one of article popularity, user age and the number of articles browsed by the user per day; the a priori discrete feature data includes at least one of article payment status and whether the user recharges; the a priori continuous feature data includes at least one of the number of article comments, the number of article clicks and the number of times users comment on articles.

[0041] In a third aspect, the present invention provides a method for predicting a product's click-through rate, the method comprising:

[0042] A product prediction model trained using the model training method described in any one of the first aspects;

[0043] In response to an order request from a target user, obtaining user characteristic information of the target user, request characteristic information of the target user, and product characteristic information of an alternative product;

[0044] The user feature information, the request feature information, and the product feature information are input into the product prediction model to predict a click probability value corresponding to each of the candidate products.

[0045] Preferably, the prediction method further includes:

[0046] The candidate products are sorted according to the click probability values.

[0047] In a fourth aspect, the present invention provides a product click-through rate prediction system, the prediction system comprising:

[0048] The model training system according to any one of the second aspects, used for training a product prediction model;

[0049] A second acquisition module is configured to acquire user characteristic information of the target user, request characteristic information of the target user, and product characteristic information of the candidate product in response to the order request of the target user;

[0050] The prediction module is used to input the user feature information, the request feature information and the product feature information into the product prediction model to predict the click probability value corresponding to each of the candidate products.

[0051] Preferably, the prediction system further includes:

[0052] A processing module is used to sort the candidate products according to the click probability values.

[0053] In a fifth aspect, the present invention also provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the model training method as described in any one of the first aspects is implemented, or the product click-through rate prediction method as described in any one of the third aspects is executed.

[0054] In a sixth aspect, the present invention further provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the computer program implements the model training method as described in any one of the first aspects, or executes the steps of the product click-through rate prediction method as described in any one of the third aspects.

[0055] The positive progress of the present invention is to provide a model training method, a product click-through rate prediction method, a system, a device, and a medium. The model training method obtains data to be trained; selects prior feature data and non-prior feature data from the feature data of the training user, the feature data of the training product, and the feature data of the training user's request; and inputs the prior feature data, non-prior feature data, and click interaction data into a first model for training to obtain a product prediction model. By adding a second neural network layer that strengthens the prior features, the present invention effectively helps the model better learn the prior features, solves the problem that the prior strongly correlated features have a poor learning effect in the product prediction model, enhances the model's prediction effect, and effectively improves the accuracy of the prediction of the probability value of the product being clicked by the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart of the model training method of embodiment 1 of the present invention.

[0057] Figure 2 This is a schematic diagram of the first model structure of the model training method of Example 1 of the present invention.

[0058] Figure 3 Schematic diagram of the second model structure of the model training method of Example 1 of the present invention.

[0059] Figure 4 This is a module diagram of the model training system of Example 2 of the present invention.

[0060] Figure 5 This is a flowchart of the method for predicting product click-through rate according to Example 3 of the present invention.

[0061] Figure 6 This is a module diagram of a product click-through rate prediction system according to Example 4 of the present invention.

[0062] Figure 7This is a schematic diagram of the hardware structure of an electronic device according to embodiment 5 of the present invention. DETAILED DESCRIPTION

[0063] The present invention is further described below by way of examples, but the present invention is not limited to the scope of the examples.

[0064] Example 1

[0065] This embodiment provides a model training method, referring to Figure 1 , the method comprises the following steps:

[0066] S1. Obtain data to be trained; the data to be trained includes characteristic data of the training user, characteristic data of the training product, characteristic data of the training user's request, and click interaction data of the training user on the training product.

[0067] S2. Filter out priori feature data and non-priori feature data from the feature data of the training user, the feature data of the training product, and the feature data of the training user's request.

[0068] S3. Input the prior feature data, non-prior feature data, and click interaction data into the first model for training to obtain a product prediction model; the first model includes a first neural network layer and a second neural network layer that strengthens the prior features. The product prediction model is used to calculate the probability value of each product being clicked by the user in future orders in a future time period.

[0069] For step S1 above, the amount of data collected for training can reach over a million. The more training data collected, the more accurate the trained product prediction model. Tracking can be performed through a website or application (app) to obtain historical user click and behavior data, thereby mining training data. Each piece of collected training data includes at least two feature data of the training user, at least two feature data of the training product, at least two request feature data of the training user, and at least two click interaction data of the training user on the training product.

[0070] In different application scenarios, the characteristic data of the training user may include the user's behavior information, the user's interest tags, and the user's demographic attributes (for example, age, gender, education level, and nature of work). The characteristic information of the training product may include the product's identification, product classification, and product price. The product's identification is used to uniquely identify the product and may use at least one or a combination of at least two of numbers, English letters, and punctuation marks. The training user's request characteristic data may include the current request time and weather conditions. The training user's click interaction data on the training product may include whether the user clicked on the training product or not.

[0071] It should be noted that the executor of the model training method of this embodiment is an electronic entity, or it can also be a software integrated application. When in use, the application is run on a large computer device, and the technical method of this embodiment is used to train the product prediction model.

[0072] Regarding step S2 above, the prior feature data represents strongly correlated features that can influence the model's prediction accuracy, while the non-prior feature data represents less strongly correlated features that have a lesser impact on the model's prediction accuracy. For example, prior feature data for training users includes whether the user is a member, whether the user has topped up their account, and the number of times the user has commented on an article; prior feature data for training products includes whether the author of the article is free, article comment data, and article click counts. It should be noted that in this embodiment, the product can be a commodity, or other information such as an advertisement.

[0073] Regarding the above step S3, before training the first model, it is necessary to randomly initialize the parameters of the neural network layer in the first model, and start training the first model based on the randomly initialized parameters. Figure 2 In the process of training the first model, it is necessary to first embed the user, product and request context based on the feature data of the training user, the feature data of the training product and the feature data of the training user's request to obtain the corresponding combined feature expression. In this embodiment, the feature expression can be specifically in the form of a vector, which is called a vector expression.

[0074] In one possible implementation, step S3 includes:

[0075] S31. After inputting the non-prior feature data into the first neural network layer and the prior feature data into the second neural network layer, the first model is trained based on the click interaction data.

[0076] With respect to step S31 above, the non-prior feature data is input into multiple conventional first neural network layers, and the prior feature data is input into the second neural network layer of the Enhance Prior Unit (EPU) network structure. Based on the combined feature expressions of the training users, the combined feature expressions of the training products, and the combined feature expressions of the training users' request feature data, a first combined feature expression corresponding to the non-prior feature data and a second combined feature expression corresponding to the prior feature data are obtained. Based on the click interaction data, the first combined feature expression, the second combined feature expression, and a pre-defined loss function, the first model is iterated, and the parameters of the first model are continuously adjusted to generate a product prediction model.

[0077] In a possible implementation, the non-a priori feature data includes discrete feature data and continuous feature data, and the a priori feature data includes a priori discrete feature data and a priori continuous feature data. Step S31 specifically includes:

[0078] Embedding technology is used to preprocess discrete feature data and prior discrete feature data.

[0079] The processed discrete feature data and continuous feature data are respectively processed by at least two layers including a connection layer and a fully connected layer to obtain a first combined feature expression.

[0080] The processed prior discrete feature data and the prior continuous feature data are respectively processed by at least two layers including a connection layer and a processing layer for strengthening the prior features to obtain a second combined feature expression.

[0081] The first combined feature expression and the second combined feature expression are spliced ​​and sequentially processed through at least three layers including a connection layer, a fully connected layer, and an activation processing layer to obtain the predicted click probability of the training user for the training product.

[0082] The binary cross entropy loss function is fitted based on the predicted click probability and the click interaction data of training users on the training products to generate a product prediction model.

[0083] Among them, discrete feature data includes at least one of article type, article ID, user gender, user occupation and weather conditions, continuous feature data includes at least one of article popularity, user age and the number of articles browsed by users per day, a priori discrete feature data includes at least one of article payment status and whether the user recharges, and a priori continuous feature data includes at least one of the number of article comments, the number of article clicks and the number of times users comment on articles.

[0084] Specifically, see Figure 3 , use Embedding technology to preprocess the discrete feature data in the embedding layer, and at the same time use Embedding technology to preprocess the prior discrete feature data in the embedding layer. Input the preprocessed discrete feature data and continuous feature data into the connection layer (Concat layer) and the fully connected layer (Dense layer) of the first neural network layer, and output the first combined feature expression. Input the preprocessed prior discrete feature data and prior continuous feature data into the second neural network layer that strengthens the prior features, and output the second combined feature expression. After splicing the first combined feature expression and the second combined feature expression, they are processed in sequence by the connection layer (Concat layer), the fully connected layer (Dense layer), and the activation processing layer (Sigmoid layer) to generate a product prediction model.

[0085] In one possible implementation, the second neural network layer for strengthening the prior features includes a first fully connected network unit, a second fully connected network unit, a Softmax activation function processing network unit, and a processing network unit; the Softmax activation function processing network unit is located between the first fully connected network unit and the processing network unit;

[0086] The input of the Softmax activation function processing network unit includes the output of the first fully connected network unit, and the output of the Softmax activation function processing network unit includes the enhancement coefficient of each feature expression of the output of the second fully connected network unit;

[0087] The input of the processing network unit includes the output of the Softmax activation function processing network unit and the output of the second fully connected network unit, and the dimension of the feature expression of the output of the Softmax activation function processing network unit and the feature expression of the output of the second fully connected network unit are the same.

[0088] Specifically, Figure 2 This is a schematic diagram of the structure of the second neural network layer for strengthening the prior features, that is, a schematic diagram of the structure of the network unit for strengthening the prior features. After the prior discrete feature data is processed by Embedding and spliced ​​with the prior continuous feature data, it is input into the EPU. All feature data are input into the x3 layer (Dense layer) on the right, and all feature data are input into the x1 layer (Dense layer) on the left. Both layers output n-dimensional vectors. After the x1 layer, a Softmax activation function is added to output an n-dimensional vector with a sum of 1 values, and 1 is added to the output result element by element, so that the element value of the final x2 layer output n-dimensional vector is between 1 and 2. The output vector of the x2 layer and the output vector of the x3 layer are subjected to a Hadamard product operation, so that the final output result is weighted by the prior feature data on the output result of the x3 layer, and the strengthening coefficient is set between [1, 2].

[0089] If the amount of feature data of training users, feature data of training products, and feature data requested by training users is large, more levels of connection layers (Concat layers) or fully connected layers (Dense layers) can be set. The specific number of connection layers and fully connected layers can be determined according to the amount of data to be trained, so that the feature data is mixed more evenly, and ultimately the prediction accuracy of the product prediction model is improved.

[0090] In this embodiment, the present invention effectively helps the model to better learn the prior features by adding a second neural network layer to strengthen the prior features, solves the problem that the prior strongly correlated features are not obvious in the product prediction model, enhances the prediction effect of the model, and effectively improves the prediction accuracy of the probability value of the product being clicked by the user.

[0091] Example 2

[0092] This embodiment provides a model training system, referring to Figure 4 , specifically including: a first acquisition module 110, a screening module 120 and a training module 130.

[0093] The first acquisition module 110 is used to acquire data to be trained; the data to be trained includes characteristic data of the training user, characteristic data of the training product, characteristic data of the training user's request, and click interaction data of the training user on the training product.

[0094] The screening module 120 is configured to screen out priori feature data and non-priori feature data from the feature data of the training user, the feature data of the training product, and the feature data requested by the training user.

[0095] The training module 130 is used to input the prior feature data and the non-prior feature data into the first model for training to obtain a product prediction model; the first model includes a first neural network layer and a second neural network layer that strengthens the prior features. The product prediction model is used to calculate the probability value of each product being clicked by the user in future orders in the future time period.

[0096] The amount of data collected for training can reach over a million. The more training data collected, the more accurate the trained product prediction model. Tracking can be performed through websites or applications (apps) to obtain historical user click and behavior data, thereby mining training data. Each piece of training data collected by the first acquisition module 110 includes at least two feature data of the training user, at least two feature data of the training product, at least two request feature data of the training user, and at least two click interaction data of the training user on the training product.

[0097] In different application scenarios, the characteristic data of the training user may include the user's behavior information, the user's interest tags, and the user's demographic attributes (for example, age, gender, education level, and nature of work). The characteristic information of the training product may include the product's identification, product classification, and product price. The product's identification is used to uniquely identify the product and may use at least one or a combination of at least two of numbers, English letters, and punctuation marks. The training user's request characteristic data may include the current request time and weather conditions. The training user's click interaction data on the training product may include whether the user clicked on the training product or not.

[0098] It should be noted that the executor of the model training method of this embodiment is an electronic entity, or it can also be a software integrated application. When in use, the application is run on a large computer device, and the technical method of this embodiment is used to train the product prediction model.

[0099] Prior feature data represents strongly correlated features that can influence the model's prediction accuracy, while non-prior feature data represents less strongly correlated features that have a lesser impact on the model's prediction accuracy. For example, prior feature data for training users includes whether the user is a member, whether the user has topped up their account, and the number of times the user has commented on an article; prior feature data for training products includes whether the author of the article is free, article comment data, and article click counts. It should be noted that in this embodiment, the product can be a commodity, or other information such as an advertisement.

[0100] Before training the first model, the parameters of the neural network layer in the first model need to be randomly initialized, and then training of the first model begins based on the randomly initialized parameters. During the training of the first model, the training module 130 needs to first embed the user, product, and request contexts based on the feature data of the training user, the feature data of the training product, and the feature data of the training user's request, respectively, to obtain the corresponding combined feature expression. In this embodiment, the feature expression can specifically be in the form of a vector, which is called a vector expression.

[0101] In one possible implementation, refer to Figure 4 , the training module 130 includes a training unit 131.

[0102] The training unit 131 is used to input the non-prior feature data into the first neural network layer and the prior feature data into the second neural network layer, and then train the first model based on the click interaction data.

[0103] The non-a priori feature data is input into multiple conventional first neural network layers, and the a priori feature data is input into the second neural network layer of the enhanced a priori feature network structure (EPU). A first combined feature expression corresponding to the non-a priori feature data and a second combined feature expression corresponding to the a priori feature data are obtained by concatenating the combined feature expressions of the training users, the combined feature expressions of the training products, and the combined feature expressions of the training users' request feature data. The training unit 131 iterates the first model based on the click interaction data, the first combined feature expression, the second combined feature expression, and a pre-defined loss function, and continuously adjusts the parameters of the first model to generate a product prediction model.

[0104] In one possible implementation, the non-a priori feature data includes discrete feature data and continuous feature data, the a priori feature data includes a priori discrete feature data and a priori continuous feature data, and the training unit 131 is specifically configured to:

[0105] Embedding technology is used to preprocess discrete feature data and prior discrete feature data.

[0106] The processed discrete feature data and continuous feature data are respectively processed by at least two layers including a connection layer and a fully connected layer to obtain a first combined feature expression.

[0107] The processed prior discrete feature data and the prior continuous feature data are respectively processed by at least two layers including a connection layer and a processing layer for strengthening the prior features to obtain a second combined feature expression.

[0108] The first combined feature expression and the second combined feature expression are spliced ​​and sequentially processed through at least three layers including a connection layer, a fully connected layer, and an activation processing layer to obtain the predicted click probability of the training user for the training product.

[0109] The binary cross entropy loss function is fitted based on the predicted click probability and the click interaction data of training users on the training products to generate a product prediction model.

[0110] Among them, discrete feature data includes at least one of article type, article ID, user gender, user occupation and weather conditions, continuous feature data includes at least one of article popularity, user age and the number of articles browsed by users per day, a priori discrete feature data includes at least one of article payment status and whether the user recharges, and a priori continuous feature data includes at least one of the number of article comments, the number of article clicks and the number of times users comment on articles.

[0111] Specifically, the Embedding technology is used to preprocess the discrete feature data in the embedding layer, and the Embedding technology is used to preprocess the prior discrete feature data in the embedding layer. The preprocessed discrete feature data and continuous feature data are input into the connection layer (Concat layer) and the fully connected layer (Dense layer) of the first neural network layer, and the first combined feature expression is output. The preprocessed prior discrete feature data and the prior continuous feature data are input into the second neural network layer that strengthens the prior features, and the second combined feature expression is output. After the first combined feature expression and the second combined feature expression are spliced ​​together, they are processed in sequence by the connection layer (Concat layer), the fully connected layer (Dense layer) and the activation processing layer (Sigmoid layer) to generate a product prediction model.

[0112] In one possible implementation, the second neural network layer for strengthening the prior features includes a first fully connected network unit, a second fully connected network unit, a Softmax activation function processing network unit, and a processing network unit; the Softmax activation function processing network unit is located between the first fully connected network unit and the processing network unit;

[0113] The input of the Softmax activation function processing network unit includes the output of the first fully connected network unit, and the output of the Softmax activation function processing network unit includes the enhancement coefficient of each feature expression of the output of the second fully connected network unit;

[0114] The input of the processing network unit includes the output of the Softmax activation function processing network unit and the output of the second fully connected network unit, and the dimension of the feature expression of the output of the Softmax activation function processing network unit and the feature expression of the output of the second fully connected network unit are the same.

[0115] Specifically, Figure 2 This is a schematic diagram of the structure of the second neural network layer for enhancing the prior features, that is, a schematic diagram of the structure of the Enhance Prior Unit (hereinafter referred to as EPU). After the prior discrete feature data is processed by Embedding and spliced ​​with the prior continuous feature data, it is input into the EPU. All feature data are input into the x3 layer (Dense layer) on the right, and all feature data are input into the x1 layer (Dense layer) on the left. Both layers output n-dimensional vectors. After the x1 layer, an n-dimensional vector with a sum of Softmax output values ​​of 1 is added, and 1 is added to the output result element by element, so that the element value of the final x2 layer output n-dimensional vector is between 1 and 2. The output vector of the x2 layer and the output vector of the x3 layer are operated with a Hadamard product, so that the final output result is weighted by the prior feature data on the output result of the x3 layer, and the enhancement coefficient is set between [1, 2].

[0116] If the amount of feature data of training users, feature data of training products, and feature data requested by training users is large, more levels of connection layers (Concat layers) or fully connected layers (Dense layers) can be set. Specifically, the number of connection layers and fully connected layers can be determined according to the amount of data to be trained, so that the feature data is mixed more evenly, and ultimately the prediction accuracy of the product prediction model is improved.

[0117] The model training system of this embodiment realizes model training by adopting the above modules, and the implementation principle and technical effects are the same as those of the above-mentioned related method embodiments, which will not be repeated here.

[0118] In this embodiment, the present invention utilizes a training module to build a first model using a second neural network layer that incorporates enhanced prior features, combined with prior feature data for training and generating a product prediction model. This effectively helps the model better learn the prior features, resolving the issue of weak learning effects of strongly correlated prior features in the product prediction model. This enhances the model's prediction performance and effectively improves the accuracy of predicting the probability of a product being clicked by a user.

[0119] Example 3

[0120] This embodiment provides a method for predicting the click-through rate of a product. Figure 5 , the prediction method includes:

[0121] S21. A product prediction model is trained using the model training method of Example 1.

[0122] S22 . In response to the order request of the target user, obtain user characteristic information of the target user, request characteristic information of the target user, and product characteristic information of the candidate product.

[0123] S23: Input the user feature information, request feature information, and product feature information into a product prediction model to predict a click probability value corresponding to each candidate product.

[0124] S24. Sort the candidate products according to the click probability values.

[0125] Specifically, when a target user requests a product, the system obtains the characteristics of recalled items that meet the requirements and determines the product feature information of the candidate products from these characteristics. Using the trained product prediction model, the system predicts the likelihood of the target user clicking on each candidate product, generating a click probability value for each candidate product. For example, user feature information includes the user's name, contact information, age, gender, and location; request feature information includes the target user's request date and weather conditions; and product feature information includes the hotel name, address, and model.

[0126] It should be noted that a larger click probability value indicates a higher target user interest in the candidate product and a higher probability of clicking on it. Conversely, a smaller click probability value indicates a lower target user interest in the candidate product and a lower probability of clicking on it. Recommending products based on the click probability values ​​corresponding to candidate products output by the product prediction model can effectively improve recommendation accuracy.

[0127] This embodiment provides a method for predicting product click-through rates, which effectively improves the accuracy of predicting the probability of a product being clicked by a user by using a trained product prediction model to predict the click probability values ​​of candidate products.

[0128] Example 4

[0129] This embodiment provides a product click rate prediction system, referring to Figure 6 The prediction system includes: the model training system of the above-mentioned embodiment 3, a second acquisition module 210, a prediction module 220 and a processing module 230.

[0130] The model training system of the above-mentioned embodiment 3 is used to train a product prediction model.

[0131] The second acquisition module 210 is configured to acquire user characteristic information of the target user, request characteristic information of the target user, and product characteristic information of the candidate products in response to the order request of the target user.

[0132] The prediction module 220 is used to input user feature information, request feature information and product feature information into the product prediction model to predict the click probability value corresponding to each candidate product.

[0133] The processing module 230 is used to sort the candidate products according to the click probability values.

[0134] Specifically, when a target user requests a product, the system obtains the characteristics of recalled items that meet the requirements and determines the product feature information of the candidate products from these characteristics. Using the trained product prediction model, the system predicts the likelihood of the target user clicking on each candidate product, generating a click probability value for each candidate product. For example, user feature information includes the user's name, contact information, age, gender, and location; request feature information includes the target user's request date and weather conditions; and product feature information includes the hotel name, address, and model.

[0135] It should be noted that a larger click probability value indicates a higher target user interest in the candidate product and a higher probability of clicking on it. Conversely, a smaller click probability value indicates a lower target user interest in the candidate product and a lower probability of clicking on it. Recommending products based on the click probability values ​​corresponding to candidate products output by the product prediction model can effectively improve recommendation accuracy.

[0136] This embodiment provides a product click-through rate prediction system, which effectively improves the accuracy of predicting the probability of a product being clicked by a user by using a trained product prediction model to predict the click probability value of an alternative product.

[0137] Example 5

[0138] Figure 7This is a schematic diagram of the structure of an electronic device provided in this embodiment. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the model training method of embodiment 1 or the product click-through rate prediction method of embodiment 3 is implemented. Figure 7 The electronic device 30 shown is only an example and should not limit the functionality and scope of use of the embodiments of the present invention.

[0139] The electronic device 30 may be a general-purpose computing device, such as a server device. Components of the electronic device 30 may include, but are not limited to, the at least one processor 31, the at least one memory 32, and a bus 33 connecting various system components (including the memory 32 and the processor 31).

[0140] The bus 33 includes a data bus, an address bus, and a control bus.

[0141] The memory 32 may include a volatile memory, such as a random access memory (RAM) 321 and / or a cache memory 322 , and may further include a read-only memory (ROM) 323 .

[0142] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, such program modules 324 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0143] The processor 31 executes various functional applications and data processing by running the computer programs stored in the memory 32, such as the model training method of Example 1 of the present invention or the product click-through rate prediction method of Example 3.

[0144] The electronic device 30 can also communicate with one or more external devices 34 (e.g., a keyboard, pointing device, etc.). This communication can occur via an input / output (I / O) interface 35. Furthermore, the model-generating device 30 can also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 36. As shown, the network adapter 36 communicates with other modules of the model-generating device 30 via a bus 33. It should be understood that, although not shown, other hardware and / or software modules can be used in conjunction with the model-generating device 30, including but not limited to microcode, device drivers, redundant processors, external disk drive arrays, RAID (RAID) systems, tape drives, and data backup storage systems.

[0145] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more units / modules described above may be embodied in a single unit / module. Conversely, the features and functions of a single unit / module described above may be further divided and embodied by multiple units / modules.

[0146] Example 6

[0147] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the model training method of embodiment 1 or the steps of the product click-through rate prediction method of embodiment 3 are implemented.

[0148] The readable storage medium may include, but is not limited to, a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0149] In a possible embodiment, the present invention can also be implemented in the form of a program product, which includes program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps of the model training method of Example 1 or the steps of the product click-through rate prediction method of Example 3.

[0150] The program code for executing the present invention may be written in any combination of one or more programming languages, and may be executed entirely on the user device, partially on the user device, as an independent software package, partially on the user device and partially on a remote device, or entirely on the remote device.

[0151] Although specific embodiments of the present invention have been described above, those skilled in the art will appreciate that these are merely illustrative and that the scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, and such changes and modifications are intended to fall within the scope of the present invention.

Claims

1. A model training method, characterized in that: include: Get the data to be trained; The data to be trained includes characteristic data of the training user, characteristic data of the training product, characteristic data of the training user's request, and click interaction data of the training user on the training product; Filtering out priori feature data and non-priori feature data from the feature data of the training user, the feature data of the training product, and the feature data of the training user's request; Inputting the prior feature data, the non-prior feature data, and the click interaction data into a first model for training to obtain a product prediction model; the first model includes a first neural network layer and a second neural network layer that strengthens the prior features, and the product prediction model is used to calculate the probability value of each product being clicked by the user in future orders in a future time period; The second neural network layer that strengthens the prior features includes a first fully connected network unit, a second fully connected network unit, a Softmax activation function processing network unit and a processing network unit; the Softmax activation function processing network unit is located between the first fully connected network unit and the processing network unit; The input of the Softmax activation function processing network unit includes the output of the first fully connected network unit, and the output of the Softmax activation function processing network unit includes the enhancement coefficient of each feature expression of the output of the second fully connected network unit; The input of the processing network unit includes the output of the Softmax activation function processing network unit and the output of the second fully connected network unit, and the dimension of the feature expression of the output of the Softmax activation function processing network unit and the feature expression of the output of the second fully connected network unit are the same.

2. The model training method according to claim 1, wherein: The step of inputting the priori feature data, the non-priori feature data, and the click interaction data into a first model for training to obtain a product prediction model includes: After inputting the non-prior feature data into the first neural network layer and the prior feature data into the second neural network layer respectively, the first model is trained based on the click interaction data.

3. The model training method according to claim 2, wherein: The non-a priori feature data includes discrete feature data and continuous feature data, and the a priori feature data includes a priori discrete feature data and a priori continuous feature data; After inputting the non-a priori feature data into the first neural network layer and the a priori feature data into the second neural network layer, the step of training the first model based on the click interaction data includes: Preprocessing the discrete feature data and the prior discrete feature data using Embedding technology; Processing the processed discrete feature data and the processed continuous feature data through at least two layers including a connection layer and a fully connected layer, respectively, to obtain a first combined feature expression; Processing the processed prior discrete feature data and the processed prior continuous feature data respectively through at least two layers including a connection layer and a processing layer for strengthening prior features to obtain a second combined feature expression; The first combined feature expression and the second combined feature expression are concatenated and sequentially processed through at least three layers including a connection layer, a fully connected layer, and an activation processing layer to obtain a predicted click probability of the training user for the training product; The binary cross entropy loss function is fitted according to the predicted click probability and the click interaction data to generate the product prediction model.

4. The model training method according to claim 3, wherein: The discrete feature data includes at least one of article type, article ID, user gender, user occupation and weather conditions; the continuous feature data includes at least one of article popularity, user age and the number of articles browsed by the user per day; the a priori discrete feature data includes at least one of article payment status and whether the user recharges; the a priori continuous feature data includes at least one of the number of article comments, the number of article clicks and the number of times users comment on articles.

5. A model training system, characterized in that: include: A first acquisition module is used to acquire data to be trained; The data to be trained includes characteristic data of the training user, characteristic data of the training product, characteristic data of the training user's request, and click interaction data of the training user on the training product; A screening module, configured to screen out priori feature data and non-priori feature data from the feature data of the training user, the feature data of the training product, and the feature data requested by the training user; a training module configured to input the prior feature data, the non-prior feature data, and the click interaction data into a first model for training to obtain a product prediction model; the first model comprising a first neural network layer and a second neural network layer that reinforces the prior features; the product prediction model being configured to calculate a probability value of each product being clicked by a user in a future order in a future time period; The second neural network layer that strengthens the prior features includes a first fully connected network unit, a second fully connected network unit, a Softmax activation function processing network unit and a processing network unit; the Softmax activation function processing network unit is located between the first fully connected network unit and the processing network unit; The input of the Softmax activation function processing network unit includes the output of the first fully connected network unit, and the output of the Softmax activation function processing network unit includes the enhancement coefficient of each feature expression of the output of the second fully connected network unit; The input of the processing network unit includes the output of the Softmax activation function processing network unit and the output of the second fully connected network unit, and the dimension of the feature expression of the output of the Softmax activation function processing network unit and the feature expression of the output of the second fully connected network unit are the same.

6. The model training system according to claim 5, wherein: The training module includes: A training unit is used to input the non-prior feature data into the first neural network layer and the prior feature data into the second neural network layer, and then train the first model based on the click interaction data.

7. The model training system according to claim 6, wherein: The non-a priori feature data includes discrete feature data and continuous feature data, and the a priori feature data includes a priori discrete feature data and a priori continuous feature data; The training unit is specifically used for: Preprocessing the discrete feature data and the prior discrete feature data using Embedding technology; Processing the processed discrete feature data and the processed continuous feature data through at least two layers including a connection layer and a fully connected layer, respectively, to obtain a first combined feature expression; Processing the processed prior discrete feature data and the processed prior continuous feature data respectively through at least two layers including a connection layer and a processing layer for strengthening prior features to obtain a second combined feature expression; The first combined feature expression and the second combined feature expression are concatenated and sequentially processed through at least three layers including a connection layer, a fully connected layer, and an activation processing layer to obtain a predicted click probability of the training user for the training product; The binary cross entropy loss function is fitted according to the predicted click probability and the click interaction data of the training users on the training product to generate the product prediction model.

8. The model training system according to claim 7, wherein: The discrete feature data includes at least one of article type, article ID, user gender, user occupation and weather conditions; the continuous feature data includes at least one of article popularity, user age and the number of articles browsed by the user per day; the a priori discrete feature data includes at least one of article payment status and whether the user recharges; the a priori continuous feature data includes at least one of the number of article comments, the number of article clicks and the number of times users comment on articles.

9. A method for predicting product click-through rate, characterized in that: The prediction method comprises: A product prediction model trained using the model training method according to any one of claims 1 to 4; In response to an order request from a target user, obtaining user characteristic information of the target user, request characteristic information of the target user, and product characteristic information of an alternative product; The user feature information, the request feature information, and the product feature information are input into the product prediction model to predict a click probability value corresponding to each of the candidate products.

10. The method for predicting product click-through rate according to claim 9, wherein: The prediction method further comprises: The candidate products are sorted according to the click probability values.

11. A product click-through rate prediction system, characterized in that: The prediction system includes: The model training system according to any one of claims 5 to 8, used for training a product prediction model; A second acquisition module is configured to acquire user characteristic information of the target user, request characteristic information of the target user, and product characteristic information of the candidate product in response to the order request of the target user; The prediction module is used to input the user feature information, the request feature information and the product feature information into the product prediction model to predict the click probability value corresponding to each of the candidate products.

12. The product click-through rate prediction system according to claim 11, wherein: The prediction system further includes: A processing module is used to sort the candidate products according to the click probability values.

13. An electronic device, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the model training method according to any one of claims 1 to 4, or executes the product click-through rate prediction method according to any one of claims 9 to 10.

14. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the model training method according to any one of claims 1 to 4, or executes the product click-through rate prediction method according to any one of claims 9 to 10.

Images