Click rate model training method, article recommendation method and equipment

Abstract

The invention provides a click rate model training method, and an article recommendation method and device.The click rate model training method comprises the steps that a vector fusion layer in an initial click rate model is adopted to fuse a first user feature, a first cooperation feature and a first article feature sequence of a sample object to obtain a first fusion vector; performing first feature processing on the first fusion vector by adopting a feature processing layer in an initial click rate model to obtain a second fusion vector; and according to the first fusion vector, the second fusion vector and the first material feature sequence of the first material for displaying the plurality of sample articles, determining weights of the plurality of sample articles, so as to train the initial click rate model according to the weights of the plurality of sample articles, thereby improving the effectiveness of training the initial click rate model, and improving the accuracy of training the initial click rate model. Therefore, the click rate of each article predicted based on the trained click rate model can accurately recommend the article to the user, and the user experience is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular, to a method for training a click-through rate model, an item recommendation method, and a device. Background Art

[0002] Currently, in an e-commerce platform, items are displayed through materials such as short videos, stores, live broadcasts, activities, and commodities, which can provide users with a more rich and interesting shopping experience and improve the sales ability of the e-commerce platform.

[0003] However, due to the unequal exposure of the above-mentioned materials, when separately modeling each material, the click-through rate model may deviate towards the sample with a larger proportion because the sample distribution in each batch of training is quite different, resulting in the trained click-through rate model being unable to accurately predict the click-through rate of each item. As a result, items cannot be accurately recommended to users based on the predicted click-through rate of each item, reducing the user experience. Summary of the Invention

[0004] The present disclosure aims to solve at least one of the technical problems in the related art to some extent.

[0005] To this end, the present disclosure proposes a method for training a click-through rate model, which can determine the weights of multiple sample items according to the first object feature, the first collaborative feature, and the first item feature sequence of the sample object, as well as the first material feature sequence of the first material that displays multiple sample items, and train an initial click-through rate model according to the weights of the multiple sample items. Thus, the display fluctuation between the materials for displaying each sample item can be considered evenly, so as to balance the distribution of the sample items for training the initial click-through rate model, reduce the probability that the click-through rate model deviates towards the sample with a larger proportion due to the large difference in sample distribution during the training of the initial click-through rate model, improve the effectiveness of training the click-through rate model, and thus improve the accuracy of the trained click-through rate model in predicting the click-through rate of each item. Based on the predicted click-through rate of each item, items can be accurately recommended to users, improving the user experience.

[0006] An embodiment of the first aspect of the present disclosure provides a method for training a click-through rate model, including: obtaining a first object feature, a first collaborative feature, and a first item feature sequence of a sample object, where the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes item features of multiple sample items associated with the sample object; using a vector fusion layer in an initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fused vector; using a feature processing layer in the initial click-through rate model to perform a first feature process on the first fused vector to obtain a second fused vector; determining weights of the multiple sample items according to the first fused vector, the second fused vector, and a first material feature sequence of a first material displaying the multiple sample items, and training the initial click-through rate model according to the weights of the multiple sample items.

[0007] The method for training a click-through rate model according to the embodiment of the present disclosure, by obtaining a first object feature, a first collaborative feature, and a first item feature sequence of a sample object, where the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes item features of multiple sample items associated with the sample object; using a vector fusion layer in an initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fused vector; using a feature processing layer in the initial click-through rate model to perform a first feature process on the first fused vector to obtain a second fused vector; determining weights of the multiple sample items according to the first fused vector, the second fused vector, and a first material feature sequence of a first material displaying the multiple sample items, and training the initial click-through rate model according to the weights of the multiple sample items. Thus, according to the first object feature, the first collaborative feature, and the first item feature sequence of the sample object, and the first material feature sequence of the first material displaying multiple sample items, weights of the multiple sample items are determined, and the initial click-through rate model is trained according to the weights of the multiple sample items, which can evenly consider the display fluctuation situations among the materials of each sample item shown, so as to evenly train the distribution of the sample items of the initial click-through rate model, reduce the probability that the click-through rate model deviates towards the sample with a larger proportion during the training of the initial click-through rate model due to a large difference in sample distribution, improve the effectiveness of training the click-through rate model, thereby improving the accuracy of predicting the click-through rate of each item by the trained click-through rate model. Based on the predicted click-through rate of each item, items can be accurately recommended to users, improving the user experience.

[0008] The second aspect of the present disclosure provides an item recommendation method, including: receiving a target request sent by a client, and obtaining a second collaborative feature and a second item feature sequence associated with the target object according to the second object feature of the target object in the target request, where the second collaborative feature is obtained by performing behavior statistics on target items associated with the historical behavior of the target object, and the second item feature sequence includes item features of multiple target items; screening out a second material for displaying the multiple target items that matches the historical behavior of the target object from a set item material library; using the click-through rate model trained by the method described in the first aspect of the above embodiments, predicting the click-through rate of the multiple target items according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material, so as to obtain the predicted click-through rate of the multiple target items; determining the recommendation scores of the multiple target items according to the predicted click-through rates of the multiple target items; and sending the multiple target items to the client according to the recommendation scores of the multiple target items, where the target sequence is used for the client to display the multiple target items according to the recommendation scores of the multiple target items.

[0009] The third aspect of the present disclosure provides a training device for a click-through rate model, including: an acquisition module, configured to acquire a first object feature, a first collaborative feature, and a first item feature sequence of a sample object, where the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes item features of multiple sample items associated with the sample object; a first fusion module, configured to fuse the first user feature, the first collaborative feature, and the first item feature sequence by using a vector fusion layer in an initial click-through rate model to obtain a first fusion vector; a first processing module, configured to perform first feature processing on the first fusion vector by using a feature processing layer in the initial click-through rate model to obtain a second fusion vector; a second processing module, configured to determine weights of the multiple sample items according to the first fusion vector, the second fusion vector, and a first material feature sequence of a first material for displaying the multiple sample items; and a training module, configured to train the initial click-through rate model according to the weights of the multiple sample items.

[0010] A fourth aspect embodiment of the present disclosure provides an item recommendation device, which is applied to a server and includes: a receiving module, configured to receive a target request sent by a client, and obtain a second collaborative feature and a second item feature sequence associated with the target object according to a second object feature of the target object in the target request, where the second collaborative feature is obtained by performing behavior statistics on target items associated with the historical behavior of the target object, and the second item feature sequence includes item features of multiple target items; a screening module, configured to screen out a second material for displaying the multiple target items that matches the historical behavior of the target object from a set item material library; a prediction module, configured to use a click-through rate model trained by the method described in the first aspect embodiment of the present disclosure to predict the click-through rates of the multiple target items according to the second object feature, the second collaborative feature, the second item feature sequence, and a second material feature sequence of the second material, so as to obtain predicted click-through rates of the multiple target items; a determination module, configured to determine a recommendation score of the multiple target items according to the predicted click-through rates of the multiple target items; a sending module, configured to send the multiple target items to the client according to the recommendation scores of the multiple target items, where the target sequence is used for the client to display the multiple target items according to the recommendation scores of the multiple target items.

[0011] A fifth aspect embodiment of the present disclosure provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the training method of the click-through rate model described in the first aspect embodiment of the present disclosure, or implements the item recommendation method described in the second aspect embodiment of the present disclosure.

[0012] A sixth aspect embodiment of the present disclosure provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the training method of the click-through rate model described in the first aspect embodiment of the present disclosure, or implements the item recommendation method described in the second aspect embodiment of the present disclosure.

[0013] A seventh aspect embodiment of the present disclosure provides a computer program product. When the instruction processor in the computer program product executes, it implements the training method of the click-through rate model described in the first aspect embodiment of the present disclosure, or implements the item recommendation method described in the second aspect embodiment of the present disclosure.

[0014] Additional aspects and advantages of the present disclosure will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present disclosure. Description of the Drawings

[0015] The above and / or additional aspects and advantages of the present disclosure will become apparent and be readily understood from the following description of embodiments in conjunction with the accompanying drawings, where:

[0016] Figure 1 It is a schematic flowchart of a method for training a click-through rate model provided by an embodiment of the present disclosure;

[0017] Figure 2 It is a schematic flowchart of another method for training a click-through rate model provided by an embodiment of the present disclosure;

[0018] Figure 3 It is a schematic flowchart of another method for training a click-through rate model provided by an embodiment of the present disclosure;

[0019] Figure 4 It is a schematic flowchart of another method for training a click-through rate model provided by an embodiment of the present disclosure;

[0020] Figure 5 It is a schematic diagram of the training principle of the initial click-through rate model provided by an embodiment of the present disclosure;

[0021] Figure 6 It is a schematic flowchart of a method for item recommendation provided by an embodiment of the present disclosure;

[0022] Figure 7 It is a schematic flowchart of a method for item recommendation provided by an embodiment of the present disclosure;

[0023] Figure 8 It is a schematic structural diagram of a device for training a click-through rate model provided by an embodiment of the present disclosure;

[0024] Figure 9 It is a schematic structural diagram of a device for item recommendation provided by an embodiment of the present disclosure;

[0025] Figure 10 It is a block diagram of an electronic device shown according to an exemplary embodiment. Detailed Description of the Embodiment

[0026] The embodiments of the present disclosure will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions from beginning to end. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present disclosure, but should not be construed as limiting the present disclosure.

[0027] It should be noted that in the technical solution of the present disclosure, in terms of the collection, gathering, updating, analysis, processing, use, transmission, storage, etc. of user personal information, it complies with the provisions of relevant laws and regulations, is used for legal purposes, and does not violate public order and good customs. Necessary measures are taken for user personal information to prevent illegal access to user personal information data, and to safeguard user personal information security, network security, and national security.

[0028] In an e-commerce platform, item recommendation is a crucial component. It can effectively solve the problem of information overload, improve the user's shopping experience, and the sales ability of the e-commerce platform. Item recommendation based on heterogeneous materials is a new display form in the rapid development of e-commerce platforms. It expands the display method of ordinary product cards, using activities, stores, short videos, live broadcasts, and ordinary products as the card display carriers, aiming to provide users with a more diverse shopping experience. Rich heterogeneous material scenarios can improve the accuracy and personalization of item recommendation, enrich recommendation strategies, enhance the user's shopping experience, and promote commodity sales.

[0029] In item recommendation based on heterogeneous materials, it is necessary to rank multiple candidate items to determine the items that users finally see, thereby affecting users' purchase behaviors and decisions. Among them, CTR (Click-Through-Rate) is an important indicator for item ranking, which reflects the degree of user interest in items and the click-through rate. In practical applications, the click-through rate of items is predicted by using a trained CTR model. However, due to the unequal exposure of the above-mentioned materials and the large differences in the behavior distributions of users themselves, it may lead to the CTR model being biased towards the user interest samples with a larger proportion when modeling each material separately because the sample distributions in each batch of training are quite different. Although increasing the model training parameters can improve the CTR accuracy to a certain extent in small-sample scenarios, as the model parameters continue to increase, the "Hadden effect" will become more obvious, and the accuracy of the model will be more severely interfered by the heterogeneous commonalities and characteristic expressions, thus affecting the overall recommendation effect. Therefore, how to effectively train the click-through rate model is particularly crucial.

[0030] In view of the above problems, the present disclosure proposes a training method for a click-through rate model, an item recommendation method, and a device.

[0031] The following describes the training method for a click-through rate model, an item recommendation method, and a device according to embodiments of the present disclosure with reference to the accompanying drawings.

[0032] Figure 1 It is a schematic flowchart of a training method for a click-through rate model provided by an embodiment of the present disclosure.

[0033] As Figure 1 shown, the training method for the click-through rate model may include the following steps:

[0034] Step 101: Obtain the first object feature, the first collaborative feature, and the first item feature sequence of the sample object.

[0035] Among them, the first collaborative feature is obtained by performing behavior statistics on the sample items associated with the historical behavior of the sample object, and the first item feature sequence includes the item features of multiple sample items associated with the sample object.

[0036] As an example, the client that sends a request to the server at a historical moment can be determined, and the object logged in to the client is used as the sample object. Among them, the first object feature of the sample object can be obtained from the request sent by the client at the historical moment. The first object feature may include information such as the age, gender, and living habits of the sample object. The first collaborative feature is obtained by performing behavior statistics on the sample items associated with the historical behavior of the sample object. For example, how many times the sample object clicks on item A, how many times the sample object browses item B, how many times the sample object purchases item C, etc. The first item feature sequence includes the item features of multiple sample items associated with the sample object. For example, the first item feature sequence includes the item features of item A, the item features of item B, and the item features of item C, etc. Among them, the multiple sample items associated with the sample object can be multiple items associated with the historical behavior of the sample object, and the item features may include attribute information such as the color, origin, and material of the item.

[0037] Step 102: Use the vector fusion layer in the initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fusion vector.

[0038] In order to make the click-through rate model predict the click-through rate of items and conform to the shopping preferences of users, further, the vector fusion layer in the initial click-through rate model can be used to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fusion vector.

[0039] Step 103: Use the feature processing layer in the initial click-through rate model to perform first feature processing on the first fusion vector to obtain a second fusion vector.

[0040] In order to enable the click-through rate model to focus on important information in the input data and improve the generalization ability of the model, in the embodiments of the present disclosure, the feature processing layer in the initial click-through rate model can be used to perform first feature processing on the first fusion vector to obtain a second fusion vector. Among them, the feature processing layer can include an FM (Factorization Machine) sub-model and a DCN (Deep&Cross Network) sub-model, that is, the FM sub-model and the DCN sub-model are respectively used to perform feature processing on the first fusion vector, and the features after feature processing are fused to obtain a second fusion vector.

[0041] Step 104, determine the weights of multiple sample items according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material showing multiple sample items, so as to train the initial click-through rate model according to the weights of the multiple sample items.

[0042] In order to balance the distribution of sample items for training the initial click-through rate model on the basis of conforming to the shopping preferences of users, so as to reduce the probability that the click-through rate model deviates towards the sample with a larger proportion due to a large difference in sample distribution during the training of the initial click-through rate model, in the embodiments of the present disclosure, the weights of multiple sample items are determined according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material showing multiple sample items, and the initial click-through rate model is trained based on the weights of the multiple sample items. Among them, the first material feature sequence includes the identification information of the first material showing multiple sample items, and the first materials of the multiple sample items can include short videos, stores, live broadcasts, activities, and commodities (commodity cards), etc.

[0043] In summary, by determining the weights of multiple sample items according to the first object feature, the first collaborative feature, and the first item feature sequence of the sample object, as well as the first material feature sequence of the first material showing multiple sample items, and training the initial click-through rate model according to the weight vector of the multiple sample items, thus, it is possible to evenly consider the display fluctuation situation among the materials showing each sample item, balance the distribution of sample items for training the initial click-through rate model, so as to reduce the probability that the click-through rate model deviates towards the sample with a larger proportion due to a large difference in sample distribution during the training of the initial click-through rate model, improve the effectiveness of training the click-through rate model, thereby improving the accuracy of the trained click-through rate model in predicting the click-through rate of each item. Based on the predicted click-through rate of each item, items can be accurately recommended to users, improving the user experience.

[0044] In order to clearly illustrate how to determine the weights of multiple sample items according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material showing multiple sample items in the above embodiments, the present disclosure proposes another training method for the click-through rate model.

[0045] Figure 2 This is a schematic flowchart of another method for training a click-through rate model provided by an embodiment of the present disclosure.

[0046] As Figure 2 shown, the method for training the click-through rate model may include the following steps:

[0047] Step 201: Obtain the first object feature, the first collaborative feature, and the first item feature sequence of the sample object.

[0048] Among them, the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes the item features of multiple sample items associated with the sample object.

[0049] Step 202: Use the vector fusion layer in the initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fusion vector.

[0050] Step 203: Use the feature processing layer in the initial click-through rate model to perform first feature processing on the first fusion vector to obtain a second fusion vector.

[0051] As an example, the feature processing layer includes an FM sub-model and a DCN sub-model. The FM sub-model is used to perform dimensionality reduction processing on the first fusion vector to obtain a first feature, and obtain multiple low-order features of the first feature, and fuse the multiple low-order features to obtain a first sub-fusion feature; the DCN sub-model is used to perform dimensionality increase processing on the first fusion vector to obtain a second feature, and obtain multiple high-order features of the second feature, and fuse the multiple high-order features to obtain a second sub-fusion feature; the first sub-fusion feature and the second sub-fusion feature are fused to obtain a second fusion feature.

[0052] Among them, it should be noted that the FM (Factorization Machine) sub-model can decompose the first fusion vector into three parts: a global bias term, a first-order linear term, and a second-order cross term. By factorizing the first fusion vector, the FM sub-model can learn the implicit relationship between the decomposed features, thereby improving the generalization ability of the model; the DCN (Deep&Cross Network) sub-model uses the cross network to learn the internal correlation between low-order features and uses the deep network to learn the non-linear relationship between high-order features.

[0053] Step 204: Concatenate the first fusion vector and the second fusion vector to generate a third fusion vector.

[0054] In order to improve the richness of information, in the embodiments of the present disclosure, the first fusion vector and the second fusion vector may be combined to generate a third fusion vector.

[0055] Step 205: Determine the weights of multiple sample items according to the third fusion vector and the first material feature sequence.

[0056] In the embodiments of the present disclosure, for the large differences in the distributions of various materials in the model training set, which lead to mutual interference during the model learning convergence process and affect the training effect of the model. Therefore, the multi-domain personalized gating network can use the heterogeneous material feature ID as the gating feature, fuse it with the third fusion vector, determine the weights of multiple sample items, and enhance the expression of the common attributes and specific attributes between materials.

[0057] Step 206: Train the initial click-through rate model according to the weights of multiple sample items.

[0058] It should be noted that the execution processes of steps 201 to 202 and step 206 can be implemented in any one of the embodiments of the present disclosure respectively. The embodiments of the present disclosure do not make any limitations in this regard and will not be elaborated further.

[0059] In summary, the first fusion vector and the second fusion vector are concatenated to generate a third fusion vector. According to the third fusion vector and the first material feature sequence, the weights of multiple sample items are determined. Thus, by using the third fusion vector with rich information and the first material feature sequence to determine the weights of multiple sample items, the expression of the common attributes and specific attributes between materials is enhanced, and the distribution of the sample items for training the initial click-through rate model can be effectively balanced, reducing the probability that the click-through rate model deviates towards the sample with a larger proportion due to the large difference in the sample distribution during the training of the initial click-through rate model.

[0060] To clearly illustrate how to determine the weights of multiple sample items according to the third fusion vector and the first material feature sequence in the above embodiments, the present disclosure proposes another training method for the click-through rate model.

[0061] Figure 3 It is a schematic flowchart of another training method for the click-through rate model provided by the embodiments of the present disclosure.

[0062] As Figure 3 shown, the training method of the click-through rate model may include the following steps:

[0063] Step 301: Obtain the first object feature, the first collaborative feature, and the first item feature sequence of the sample object.

[0064] Among them, the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes the item features of multiple sample items associated with the sample object.

[0065] Step 302: Use the vector fusion layer in the initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fusion vector.

[0066] Step 303: Use the feature processing layer in the initial click-through rate model to perform first feature processing on the first fusion vector to obtain a second fusion vector.

[0067] Step 304: Concatenate the first fusion vector and the second fusion vector to generate a third fusion vector.

[0068] Step 305: Use the first feature extraction layer in the initial click-through rate model to extract features from the third fusion vector to obtain extracted features.

[0069] In order to further obtain important information in the third fusion vector, in the embodiments of the present disclosure, the first feature extraction layer in the initial click-through rate model can be used to extract features from the third fusion vector to obtain extracted features.

[0070] Step 306: Use the gated feature fusion layer in the initial click-through rate model to concatenate the extracted features with the first material feature sequence to obtain a first concatenated feature.

[0071] Furthermore, the gated feature fusion layer in the initial click-through rate model can be used to combine the extracted features with the first material feature sequence to obtain a first concatenated feature.

[0072] Step 307: Use the heterogeneous network gating layer in the initial click-through rate model to perform second feature processing on the first concatenated feature to obtain the weights corresponding to each sample item among multiple sample items.

[0073] As an example, the heterogeneous network gating layer includes a first multi-layer perceptron network. The first multi-layer perceptron network is used to classify the first concatenated feature to obtain the unit features of each sample item among multiple sample items; the unit feature values of each unit feature are determined through an activation function, and the unit feature values are normalized to obtain the normalized unit feature values corresponding to each unit feature value; for any sample item, the first material feature corresponding to the any sample item in the first material feature sequence is reversely activated by using the normalized unit feature value corresponding to the any sample item to obtain a weight vector corresponding to the any sample item.

[0074] Among them, it should be noted that in order to further improve the accuracy of the weight vector corresponding to any sample item, the weight matrix associated with each sample item can be multiplied by the corresponding weight vector to obtain the final weight vector of any sample item. Among them, the weight matrix associated with each sample item can be the weight matrix set according to the actual click-through rate of each sample item.

[0075] Step 308: Train the initial click-through rate model according to the weights of multiple sample items.

[0076] It should be noted that the execution processes of steps 301 to 304 and step 308 can be implemented in any one of the embodiments of the present disclosure respectively. The embodiments of the present disclosure do not make any limitations in this regard and will not be elaborated further.

[0077] In summary, by using the first feature extraction layer in the initial click-through rate model to extract features from the third fusion vector to obtain extracted features; using the gated feature fusion layer in the initial click-through rate model to splice the extracted features with the first material feature sequence to obtain the first spliced feature; using the heterogeneous network gated layer in the initial click-through rate model to perform second feature processing on the first spliced feature to obtain the weight vectors corresponding to each sample item among multiple sample items. Thus, by using the gated feature fusion layer to splice the extracted features with the first material feature sequence to obtain the first spliced feature, the expression of the common attributes and unique attributes among the first materials can be enhanced. Therefore, by using the heterogeneous network gated layer to perform second feature processing on the first spliced feature, the weight vectors of each sample item can be effectively generated.

[0078] To clearly illustrate how the above embodiments train the initial click-through rate model according to the weight vectors of multiple sample items, the present disclosure proposes another training method for the click-through rate model.

[0079] Figure 4 It is a schematic flowchart of another training method for the click-through rate model provided by the embodiments of the present disclosure.

[0080] As Figure 4 shown, the training method for the click-through rate model may include the following steps:

[0081] Step 401: Obtain the first object feature, the first collaborative feature, and the first item feature sequence of the sample object.

[0082] Among them, the first collaborative feature is obtained by statistically analyzing the sample items associated with the historical behavior of the sample object, and the first item feature sequence includes the item features of multiple sample items associated with the sample object.

[0083] Step 402: Use the vector fusion layer in the initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fusion vector.

[0084] Step 403: Use the feature processing layer in the initial click-through rate model to perform a first feature process on the first fusion vector to obtain a second fusion vector.

[0085] Step 404: Determine the weights of multiple sample items according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material showing multiple sample items.

[0086] Step 405: Obtain the first historical behavior sequence within the first set historical period associated with the sample object.

[0087] In the embodiments of the present disclosure, the first historical behavior sequence may include item information browsed, clicked, or purchased by the sample object within the first set historical period, etc., and the first historical behavior sequence can be determined according to the historical behavior of the sample object within the first set period.

[0088] Step 406: Concatenate the second fusion vector with the first historical behavior sequence to obtain a second concatenated feature.

[0089] In order to accurately predict the shopping preferences of users and improve the accuracy of item recommendation, in the embodiments of the present disclosure, a fully connected layer may be used to concatenate the second fusion vector with the first historical behavior sequence to obtain a second concatenated feature. It should be noted that, in order to capture the shopping preferences of users, before concatenating the second fusion vector with the first historical behavior sequence, the first historical behavior sequence may be serially modeled, and the modeled first historical behavior sequence is concatenated with the second fusion vector.

[0090] Step 407: Use at least one second feature extraction layer to perform feature extraction on the second concatenated feature to obtain a fourth fusion vector output by the last feature extraction layer in at least one feature extraction layer.

[0091] Furthermore, in order to obtain important features in the second concatenated feature to improve the accuracy and robustness of the click-through rate model, at least one second feature extraction layer is used to perform feature extraction on the second concatenated feature, so as to obtain a fourth fusion vector output by the last feature extraction layer in at least one feature extraction layer.

[0092] Step 408: Weight the fourth fusion vector with the weights of each sample item in multiple sample items respectively to obtain the target features of each sample item.

[0093] As an example, for any sample item, the weight vector of any sample item can be subjected to a Hadamard product with the fourth fusion vector to obtain the target feature of any sample item.

[0094] Step 409: For any sample item, sequentially input the target feature of any sample item into the corresponding second multi-layer perceptron, self-attention layer, and adaptive layer to obtain the predicted click-through rate of any sample item output by the corresponding adaptive layer.

[0095] Furthermore, sequentially input the target feature of any sample item into the corresponding second multi-layer perceptron, self-attention layer, and adaptive layer to obtain the predicted click-through rate of this sample item output by the corresponding adaptive layer.

[0096] It should be noted that the weight of the embedding vector unit under a specific material is strengthened by using the self-attention layer (Attention network), and the adaptive layer can be a Softmax function, and finally the CTR prediction accuracy of the items corresponding to each material is obtained. This technical solution can not only filter out the pure unit information under the current material and eliminate the neural information interference caused by the uneven distribution of material samples in other scenarios. At the same time, a deep neural network and an activation unit module are introduced to further amplify the model distribution prediction ability in a specific scenario.

[0097] Step 410: Train the initial click-through rate model according to the predicted click-through rate and the true click-through rate of any sample item.

[0098] As an example, according to the difference between the true click-through rate and the predicted click-through rate of any sample item, construct a maximum likelihood function; use maximum likelihood estimation to solve the maximum likelihood function to obtain the target parameter; use the target parameter as the model parameter of the trained click-through rate model.

[0099] Based on any embodiment of the present disclosure, the training principle of the initial click-through rate model in the embodiments of the present disclosure can be as Figure 5 shown, mainly including the following four aspects:

[0100] 1. Feature preparation: In the embodiments of the present disclosure, the first object feature, the first collaborative feature, and the first item feature sequence of the sample object can be obtained, as well as the first material feature (heterogeneous material feature) sequence of the first material showing multiple sample items and the first historical behavior sequence (serialized feature) within the first set historical period associated with the sample object.

[0101] 2. Multi-domain personalized gating

[0102] Since there are significant differences in the distribution of each piece of material in the model training set, which leads to interference during the model learning convergence process and affects the training effect of the model. Therefore, the vector fusion layer in the initial click-through rate model can be used to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain the first fusion vector.

[0103] For example, a multi-domain personalized gating network can be used to take the first material feature ID as the gating feature and fuse and train different material training sets to enhance the expression of common and unique attributes among materials, and achieve the transfer and expression of end-to-end user interests between different materials.

[0104] In specific implementation, the ID that can distinguish different materials (the identification information of the material) is extracted from the item feature and used as the main embedding vector alone to strengthen the expression of materials in different material scenarios. The FM output vector and the DCN output vector in the initial click-through rate model are shrunk into an embedding vector Global containing global information through a feature extractor. At the same time, the parameters of this vector are fixed by taking derivatives to prevent it from affecting the derivative convergence of the shared variables in the main tower. Next, the ReLU function is applied to the parameters in Global for screening, giving this feature a non-linear function relationship, thereby improving the fitting ability of the network.

[0105] Furthermore, the feature processing layer in the initial click-through rate model is used to perform the first feature processing on the first fusion vector to obtain the second fusion vector. Based on the first fusion vector, the second fusion vector, and the first material feature sequence of the first material that displays multiple sample items, the weights of the multiple sample items are determined.

[0106] As an example, the first fusion vector and the second fusion vector are concatenated to generate a third fusion vector. The first feature extraction layer in the initial click-through rate model is used to extract features from the third fusion vector to obtain the extracted features. The gating feature fusion layer in the initial click-through rate model is used to concatenate the extracted features with the first material feature sequence to obtain the first concatenated feature. The heterogeneous network gating layer in the initial click-through rate model is used to perform the second feature processing on the first concatenated feature to obtain the weights corresponding to each sample item among the multiple sample items. For example, a first multi-layer perceptron is used to classify the first concatenated feature to obtain the unit features of each sample item among the multiple sample items. The unit feature values of each unit feature are determined through an activation function, and the unit feature values are normalized to obtain the normalized unit feature values corresponding to each unit feature value. For any sample item, the normalized unit feature value corresponding to the sample item is used to reversely activate the first material feature corresponding to the sample item in the first material feature sequence to obtain the weight vector corresponding to the sample item.

[0107] For example, the above first fusion vector and second fusion vector are concatenated to obtain a third fusion vector. The third fusion vector passes through a layer of MLP (the first multi-layer perceptron network) with a bias vector, and a unit eigenvalue between -1 and 1 is obtained through the Tanh function (activation function). These eigenvalues are then normalized to between 0 and 1, and the ID feature vector (the first material feature corresponding to any sample item) is reversely activated through these feature units (unit eigenvalues) between 0 and 1 to obtain the weights between various materials in the current heterogeneous scenario. When finally outputting, a weight matrix assigned according to experience is multiplied according to a predetermined relationship, and then the weight matrix associated with each sample item is multiplied by the corresponding weight vector to obtain the final weight vector of any sample item. Among them, the weight matrix assigned according to experience can be the weight matrix set according to the actual click-through rate of each sample item. In this way, the common associated attributes between different materials can be obtained, and at the same time, the expression of unique attributes can be enhanced. As can be seen from the figure, the multi-domain personalized gating network can ultimately serve the multi-tower output modeling in multiple scenarios.

[0108] 3. Multi-material Adaptive Sensing Layer

[0109] For any sample item, the target features of any sample item are sequentially input into the corresponding second multi-layer perceptron network, self-attention layer, and adaptive layer to obtain the predicted click-through rate of any sample item output by the corresponding adaptive layer.

[0110] It should be understood that since the simple multi-tower modeling method can only make a relatively superficial distinction between different materials in the same scenario and does not perform fine-grained modeling on the user distribution under different materials. Therefore, in order to further distinguish the CTR prediction accuracy between different materials, the original multi-tower single-layer MLP output structure has been upgraded. For details, see Figure 5 the heterogeneous commodity material output module at the top in

[0111] For example, the multi-tower module in the original different material scenarios has been upgraded in terms of the layer of MLP (the second multi-layer perceptron network). The output of the above multi-domain personalized gating is subjected to a Hadamard product with the vector in the main tower, and the Attention network (self-attention layer) is used to strengthen the weight of the embedding vector unit under specific materials. The adaptive layer is the Softmax function, and finally the CTR prediction accuracy of each material is obtained. This technical solution can filter out the pure unit information under the current material and eliminate the neural information interference caused by the uneven distribution of material samples in other scenarios. At the same time, a deep neural network and an activation unit module are also introduced to further amplify the model distribution prediction ability in a specific scenario.

[0112] 4. Multi-domain Adaptive Loss Function

[0113] Train the initial click-through rate model according to the predicted click-through rate and the true click-through rate of any sample item. As an example, construct a maximum likelihood function based on the difference between the true click-through rate and the predicted click-through rate of any sample item; use maximum likelihood estimation to solve the maximum likelihood function to obtain the target parameters; use the target parameters as the model parameters of the trained click-through rate model.

[0114] It should be noted that in multi-material mixed learning, data distribution differences of each material are usually faced. Training methods in related technologies often only focus on the material with the largest sample size and cannot fully utilize the information of other material samples, resulting in a decline in the performance of the model on materials with a small sample size. Therefore, as an example, in order to better utilize the information of different materials and improve the generalization and expression of the model, when training the click-through rate model, an adaptive learning multi-task loss function, that is, the uncertainty loss (Uncertainty Weight Loss, abbreviated as URL), can be introduced. It can automatically learn the uncertainty of the task (that is, uncertainty), give small weights to tasks with large proportions and uncertainties, and give large weights to tasks with small proportions and large certainties, so as to adjust the convergence parameters of each material model, thereby improving the generalization ability of the model. The above loss function can be specifically expressed as the following formula:

[0115] p(y|f W (x))=Softmax(f W x);

[0116] Among them, W represents the parameter matrix of the model, f W (x) is the predicted click-through rate of each sample item, x is the input information (the first user feature, the first collaborative feature, the first item feature sequence, the first material feature sequence, and the first historical behavior sequence) when the click-through rate model predicts the click-through rate of each sample item. This representation form can be regarded as the Boltzmann distribution, and the σ 2 can be regarded as kT, that is, the product of the Boltzmann constant k and the thermodynamic temperature T. σ 2 can be a set value. According to the Boltzmann distribution, the maximum likelihood function of the above formula is:

[0117]

[0118] Among them, c represents the c-th tower in the initial click-through rate model, and c′ represents the first material feature corresponding to the c-th tower.

[0119] Furthermore, use maximum likelihood estimation to solve the maximum likelihood function to obtain the parameter matrix W of the click-through rate model.

[0120] Based on the above embodiments, the present disclosure proposes an item recommendation method.

[0121] Figure 6 Schematic flowchart of an item recommendation method provided by an embodiment of the present disclosure.

[0122] As Figure 6 shown, the item recommendation method may include the following steps:

[0123] Step 601: Receive a target request sent by a client, and obtain a second collaborative feature and a second item feature sequence associated with the target object according to the second object feature of the target object in the target request.

[0124] Among them, the second collaborative feature is obtained by performing behavior statistics on target items associated with the historical behavior of the target object, and the second item feature sequence includes item features of multiple target items.

[0125] In an embodiment of the present disclosure, the second object feature may include information such as the age, gender, and hobbies of the target object. The second collaborative feature is obtained by performing behavior statistics on target items associated with the historical behavior of the target object, and the second item feature sequence includes item features of multiple target items associated with the target object.

[0126] Step 602: Screen out a second material for displaying multiple target items that matches the historical behavior of the target object from a set item material library.

[0127] As an example, according to the historical behavior of the target object, screen out multiple second materials that meet the hobbies of the target object from the set item material library, where each second material can be used to display the corresponding target item.

[0128] Step 603: Use a trained click-through rate model to predict the click-through rates of multiple target items according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material, so as to obtain the predicted click-through rates of multiple target items.

[0129] As an example, obtain a second historical behavior sequence within a second set historical period associated with the target object; use a trained click-through rate model to predict the click-through rates of multiple target items according to the second object feature, the second collaborative feature, the second item feature sequence, the second material feature sequence, and the second historical behavior sequence, so as to obtain the predicted click-through rates of multiple target items.

[0130] That is to say, input the second object feature, the second collaborative feature, the second item feature sequence, the second material feature sequence, and the second historical behavior sequence into a trained click-through rate model, so that the click-through rate model trained in the above embodiment outputs the predicted click-through rates of each target item.

[0131] Step 604: Determine the recommendation scores of multiple target items according to the predicted click-through rates of the multiple target items.

[0132] As an example, the predicted click-through rate of each target item can be used as the recommendation score of the corresponding target item.

[0133] As another example, there may be a positive correlation between the predicted click-through rate of a target item and its corresponding recommendation score.

[0134] Step 605: Send the multiple target items to the client according to the recommendation scores of the multiple target items.

[0135] The target sequence is used by the client to display the multiple target items according to the recommendation scores of the multiple target items.

[0136] As an example, the server sorts the multiple target items according to the recommendation scores of the multiple target items to obtain a sorted sequence, and sends the sorted sequence to the client, and the client displays the multiple target items according to the sorted sequence.

[0137] As another example, the server sends the multiple target items and the scores of the multiple target items to the client, and the client sorts the multiple target items according to the recommendation scores of the multiple target items to obtain a sorted sequence, and displays the multiple target items according to the sorted sequence.

[0138] In summary, by receiving the target request sent by the client, and according to the second object feature of the target object in the target request, obtaining the second collaborative feature and the second item feature sequence associated with the target object; screening out the second materials for displaying multiple target items that match the historical behavior of the target object from the set item material library; using the trained click-through rate model, according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material, predicting the click-through rates of the multiple target items to obtain the predicted click-through rates of the multiple target items; determining the recommendation scores of the multiple target items according to the predicted click-through rates of the multiple target items; sending the multiple target items to the client according to the recommendation scores of the multiple target items. Thus, using the trained click-through rate model to predict the click-through rates of each target item improves the accuracy of determining the click-through rates of each target item, so that items can be accurately recommended to users based on the predicted click-through rates of each item, improving the user experience.

[0139] Based on any embodiment of the present disclosure, as Figure 7 shown, the item recommendation method of the embodiment of the present disclosure can also be implemented based on the following steps:

[0140] 1. System Request Trigger: The user views the recommendation page through an application (APP) or terminal and browses or clicks. These browsing and clicking behaviors of the user are received by the recommendation system service, and a request trigger instruction is sent to the system material library;

[0141] 2. Material Library Screening: The system material library covers the material (material) content to be recommended, including short videos, stores, live broadcasts, activities, products (product cards), etc. The recommendation system will parse the user's HTTP request, collect materials that match the user's interests and hobbies according to different materials, and then filter and screen them into the online recommendation module;

[0142] 3. Data Online Service: When the model is deployed online, it is necessary to analyze the existing user request data and input it into the data online service system. At the same time, it is also necessary to collect and aggregate historical data, which includes task portrait data information, such as gender and age, which can facilitate the recommendation system to make judgments. The data online service also depends on the collection and collation of historical data, including the sorting results presented by the recommendation system to users;

[0143] 4. Data Offline Service: The data offline service mainly performs operations such as filtering and sampling on the offline data of the heterogeneous material mixing model to meet the needs of data training. It includes the feature processing operations of different materials from the system material library. Through the data offline service, these data and the user's historical click behaviors are encoded in a distributed serialization manner and sent to the distributed cluster for training;

[0144] 5. Model Construction: In order to meet the huge user service requests online, we need to perform graph segmentation operations on the constructed heterogeneous sorting CTR model to meet the needs of simultaneous training on multiple machines, speed up the training speed and improve the model service inference ability. Among them, we need to allocate a heterogeneous sorting model that can calculate partial derivatives to each training machine, and all models distribute and share parameters through the central parameter sharing host to obtain uniform training of offline data;

[0145] 6. Model Push: The trained model is regularly pushed to the specified service path to meet the user's changing interests and hobbies in real time. The pushed model will be uniformly scheduled by the service scheduling system;

[0146] 7. Online Recommendation: The service scheduling system will infer the heterogeneous model and the existing data, and comprehensively score and sort the items corresponding to the heterogeneous material set that meets the user's general interest preferences. The sorting result is the user recommendation result. Finally, it is processed by the system back-end link and processed into a list of items actually recommended to the user, and the list of items is re-displayed to the user through the recommendation system service.

[0147] To achieve the above Figures 1 to 5 embodiment, the present disclosure also proposes a training device for a click-through rate model.

[0148] Figure 8 FIG. is a schematic structural diagram of a training device for a click-through rate model provided by an embodiment of the present disclosure.

[0149] As Figure 8 shown, the training device 800 for the click-through rate model includes: an acquisition module 810, a first fusion module 820, a first processing module 830, a second processing module 840, and a training module 850.

[0150] Among them, the acquisition module 810 is configured to acquire a first object feature, a first collaborative feature, and a first item feature sequence of a sample object, where the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes item features of multiple sample items associated with the sample object; the first fusion module 820 is configured to fuse the first user feature, the first collaborative feature, and the first item feature sequence by using a vector fusion layer in the initial click-through rate model to obtain a first fusion vector; the first processing module 830 is configured to perform a first feature process on the first fusion vector by using a feature processing layer in the initial click-through rate model to obtain a second fusion vector; the second processing module 840 is configured to determine weights of multiple sample items according to the first fusion vector, the second fusion vector, and a first material feature sequence of a first material that displays multiple sample items; the training module 850 is configured to train the initial click-through rate model according to the weight vectors of multiple sample items.

[0151] As a possible implementation manner of an embodiment of the present disclosure, the second processing module 840 is configured to splice the first fusion vector and the second fusion vector to generate a third fusion vector; and determine weights of multiple sample items according to the third fusion vector and the first material feature sequence.

[0152] As a possible implementation manner of an embodiment of the present disclosure, the second processing module 840 is further configured to perform feature extraction on the third fusion vector by using a first feature extraction layer in the initial click-through rate model to obtain extracted features; use a gated feature fusion layer in the initial click-through rate model to splice the extracted features with the first material feature sequence to obtain a first spliced feature; and perform a second feature process on the first spliced feature by using a heterogeneous network gating layer in the initial click-through rate model to obtain weights corresponding to each sample item among multiple sample items.

[0153] As a possible implementation manner of an embodiment of the present disclosure, the heterogeneous network gating layer includes a first multi-layer perceptron network, and the second processing module 840 is further configured to classify the first spliced feature by using the first multi-layer perceptron network to obtain the unit features of each sample item among the multiple sample items; determine the unit feature values of each unit feature through an activation function, and normalize each unit feature value to obtain the normalized unit feature value corresponding to each unit feature value; for any sample item, use the normalized unit feature value corresponding to the sample item to reversely activate the first material feature corresponding to the sample item in the first material feature sequence to obtain the weight vector corresponding to the sample item.

[0154] As a possible implementation manner of an embodiment of the present disclosure, the training module 850 is configured to obtain a first historical behavior sequence within a first set historical period associated with the sample object; splice the second fusion vector with the first historical behavior sequence to obtain a second spliced feature; perform feature extraction on the second spliced feature by using at least one second feature extraction layer to obtain a fourth fusion vector output by the last feature extraction layer in the at least one feature extraction layer; weight the fourth fusion vector with the weight vectors of each sample item among the multiple sample items to obtain the target feature of each sample item; for any sample item, sequentially input the target feature of the sample item into the corresponding second multi-layer perceptron network, self-attention layer, and adaptive layer to obtain the predicted click-through rate of the sample item output by the corresponding adaptive layer; train the initial click-through rate model according to the predicted click-through rate and the true click-through rate of any sample item.

[0155] As a possible implementation manner of an embodiment of the present disclosure, the training module 850 is further configured to construct a maximum likelihood function according to the difference between the true click-through rate and the predicted click-through rate of any sample item; solve the maximum likelihood function by using maximum likelihood estimation to obtain target parameters; use the target parameters as the model parameters of the trained click-through rate model.

[0156] As a possible implementation manner of an embodiment of the present disclosure, the feature processing layer includes an FM sub-model and a DCN sub-model, and the first processing module 830 is configured to perform dimensionality reduction processing on the first fusion vector by using the FM sub-model to obtain a first feature, obtain multiple low-order features of the first feature, and fuse the multiple low-order features to obtain a first sub-fusion feature; perform dimensionality increase processing on the first fusion vector by using the DCN sub-model to obtain a second feature, obtain multiple high-order features of the second feature, and fuse the multiple high-order features to obtain a second sub-fusion feature; fuse the first sub-fusion feature and the second sub-fusion feature to obtain the second fusion feature.

[0157] The training device for the click-through rate model according to the embodiments of the present disclosure determines the weights of multiple sample items by using the first object feature, the first collaborative feature, and the first item feature sequence of the sample object, and the first material feature sequence of the first material for displaying multiple sample items, and trains the initial click-through rate model according to the weight vectors of the multiple sample items. Thus, it is possible to evenly consider the display fluctuations between the materials for displaying each sample item, evenly train the distribution of the sample items of the initial click-through rate model, so as to reduce the probability that the click-through rate model deviates towards the sample with a larger proportion during the training of the initial click-through rate model due to a large difference in the sample distribution, improve the effectiveness of training the click-through rate model, and thus improve the accuracy of the trained click-through rate model in predicting the click-through rates of various items. Based on the predicted click-through rates of various items, items can be accurately recommended to users, improving the user experience.

[0158] To implement the above Figures 6 to 7 illustrated embodiments, the present disclosure also proposes an item recommendation device. The item recommendation device can be applied to the server side.

[0159] Figure 9 The structural schematic diagram of an item recommendation device provided by the embodiments of the present disclosure.

[0160] As Figure 9 shown, the item recommendation device 900 includes: a receiving module 910, a screening module 920, a prediction module 930, a determination module 940, and a sending module 950.

[0161] Among them, the receiving module 910 is configured to receive a target request sent by the client, and obtain a second collaborative feature and a second item feature sequence associated with the target object according to the second object feature of the target object in the target request, where the second collaborative feature is obtained by performing behavior statistics on the target items associated with the historical behavior of the target object, and the second item feature sequence includes the item features of multiple target items; the screening module 920 is configured to screen out a second material for displaying multiple target items that matches the historical behavior of the target object from a set item material library; the prediction module 930 is configured to predict the click-through rates of multiple target items through the click-through rate model obtained by training according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material, so as to obtain the predicted click-through rates of the multiple target items; where the click-through rate model is trained according to Figures 1 to 5 any one of the methods described in the embodiments in the implementation; the determination module 940 is configured to determine the recommendation scores of multiple target items according to the predicted click-through rates of the multiple target items; the sending module 950 is configured to send multiple target items to the client according to the recommendation scores of multiple target items, where the target sequence is used for the client to display multiple target items according to the recommendation scores of the multiple target items.

[0162] As a possible implementation manner of the embodiments of the present disclosure, the prediction module 930 is configured to obtain a second historical behavior sequence within a second set historical period associated with the target object; input the second object feature, the second collaborative feature, the second item feature sequence, the second material feature sequence, and the second historical behavior sequence into the trained click-through rate model to obtain the predicted click-through rates of multiple target items output by the trained click-through rate model.

[0163] The item recommendation device according to the embodiments of the present disclosure receives a target request sent by a client, and according to the second object feature of the target object in the target request, obtains a second collaborative feature and a second item feature sequence associated with the target object; filters out a second material that matches the historical behavior of the target object from a set item material library to display multiple target items; uses the trained click-through rate model to predict the click-through rates of multiple target items according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material, so as to obtain the predicted click-through rates of multiple target items; determines the recommendation scores of multiple target items according to the predicted click-through rates of multiple target items; and sends multiple target items to the client according to the recommendation scores of multiple target items. Thus, by using the trained click-through rate model to predict the click-through rates of each target item, the accuracy of determining the click-through rates of each target item is improved, and thus items can be accurately recommended to users based on the predicted click-through rates of each item, improving the user experience.

[0164] It should be noted that the foregoing explanation of the embodiments of the item recommendation method is also applicable to the item recommendation device of this embodiment, and will not be repeated here.

[0165] To implement the above embodiments, the present application also proposes an electronic device, as Figure 10 shown, Figure 10 is a block diagram of an electronic device for training a click-through rate model or item recommendation shown according to an exemplary embodiment.

[0166] As Figure 10 shown, the above-mentioned electronic device 1000 includes:

[0167] A memory 1010 and a processor 1020, a bus 1030 connecting different components (including the memory 1010 and the processor 1020), and the memory 1010 stores a computer program, and when the processor 1020 executes the program, it implements the click-through rate model training method or the item recommendation method described in the embodiments of the present disclosure.

[0168] The bus 1030 represents one or more of several types of bus architectures, including a memory bus or memory controller, a peripheral bus, an Accelerated Graphics Port, a processor bus, or a local bus using any of a variety of bus architectures. By way of example, and not limitation, these architectures include Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MAC) buses, Enhanced ISA buses, Video Electronics Standards Association (VESA) local buses, and Peripheral Component Interconnect (PCI) buses.

[0169] The electronic device 1000 typically includes a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device 1000, including both volatile and nonvolatile media, removable and non-removable media.

[0170] The memory 1010 may also include computer system-readable media in the form of volatile memory, such as random access memory (RAM) 1040 and / or cache memory 1050. The electronic device 1000 may further include other removable / non-removable, volatile / nonvolatile computer system storage media. By way of example only, a storage system 1060 can be used for reading from and writing to non-removable, nonvolatile magnetic media ( Figure 10 not shown and typically called a "hard disk drive"). Although Figure 10 not shown in the figures, a disk drive for reading from and writing to a removable nonvolatile disk (e.g., a "floppy disk"), and an optical disk drive for reading from and writing to a removable nonvolatile optical disk (e.g., a CD-ROM, a DVD-ROM, or other optical media) can be provided. In these cases, each drive can be connected to the bus 1030 by one or more data media interfaces. The memory 1010 may include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the present disclosure.

[0171] A program / utility 1080 having a set (at least one) of program modules 1070 can be stored, for example, in the memory 1010. Such program modules 1070 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data, each of which examples or some combination thereof may include an implementation of a network environment. The program modules 1070 generally carry out the functions and / or methods of the embodiments described herein.

[0172] The electronic device 1000 can also communicate with one or more external devices 1090 (such as keyboards, pointing devices, displays, etc.), and can also communicate with one or more devices that enable users to interact with the electronic device 1000, and / or communicate with any device that enables the electronic device 1000 to communicate with one or more other computing devices (such as network cards, modems, etc.). Such communication can be carried out through the input / output (I / O) interface 1092. Moreover, the electronic device 1000 can also communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) through the network adapter 1093. As Figure 10 shown, the network adapter 1093 communicates with other modules of the electronic device 1000 through the bus 1030. It should be understood that although Figure 10 not shown in the figure, other hardware and / or software modules can be used in combination with the electronic device 1000, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0173] The processor 1020 executes various functional applications and data processing by running programs stored in the memory 1010.

[0174] It should be noted that for the implementation process and technical principle of the electronic device in this embodiment, refer to the foregoing explanation of the training method of the click-through rate model or item recommendation in the embodiments of the present disclosure, which will not be elaborated here.

[0175] To implement the above embodiments, the present disclosure also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the training method of the click-through rate model or the item recommendation method described in the above embodiments.

[0176] To implement the above embodiments, the present disclosure also provides a computer program product. When the instructions in the computer program product are executed by a processor, it executes the training method of the click-through rate model or the item recommendation method described in the above embodiments.

[0177] In the description of this specification, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of this application, the meaning of "a plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0178] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0179] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A training method for a click-through rate model, characterized in that, Including: Obtain the first object feature, the first collaborative feature, and the first item feature sequence of the sample object. Among them, the first collaborative feature is obtained by performing behavior statistics on the sample items associated with the historical behavior of the sample object, and the first item feature sequence includes the item features of multiple sample items associated with the sample object; Use the vector fusion layer in the initial click-through rate model to fuse the first user feature, the first collaborative feature, and the first item feature sequence to obtain a first fusion vector; Use the feature processing layer in the initial click-through rate model to perform first feature processing on the first fusion vector to obtain a second fusion vector; Determine the weights of the multiple sample items according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material displaying the multiple sample items, and train the initial click-through rate model according to the weights of the multiple sample items.

2. The method according to claim 1, characterized in that, The determining the weights of the multiple sample items according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material displaying the multiple sample items includes: Concatenate the first fusion vector and the second fusion vector to generate a third fusion vector; Determine the weights of the multiple sample items according to the third fusion vector and the first material feature sequence.

3. The method according to claim 2, wherein The determining the weights of the multiple sample items according to the third fusion vector and the first material feature sequence includes: Use the first feature extraction layer in the initial click-through rate model to extract features from the third fusion vector to obtain extracted features; Use the gated feature fusion layer in the initial click-through rate model to concatenate the extracted features with the first material feature sequence to obtain a first concatenated feature; Use the heterogeneous network gating layer in the initial click-through rate model to perform second feature processing on the first concatenated feature to obtain the weights corresponding to each sample item among the multiple sample items.

4. The method according to claim 3, characterized in that, The heterogeneous network gating layer includes a first multi-layer perceptron. The using the heterogeneous network gating layer in the initial click-through rate model to perform second feature processing on the first concatenated feature to obtain the weights corresponding to each sample item among the multiple sample items includes: Use the first multi-layer perceptron to classify the first concatenated feature to obtain the unit features of each sample item among the multiple sample items; Determine the unit feature values of each unit feature through an activation function, and normalize each unit feature value to obtain the normalized unit feature values corresponding to each unit feature value; For any sample item, use the normalized unit feature value corresponding to the any sample item to reversely activate the first material feature corresponding to the any sample item in the first material feature sequence to obtain a weight vector corresponding to the any sample item.

5. The method according to claim 1, characterized in that, The training the initial click-through rate model according to the weights of the multiple sample items includes: Obtain the first historical behavior sequence within the first set historical period associated with the sample object; Concatenate the second fusion vector with the first historical behavior sequence to obtain a second concatenated feature; Use at least one second feature extraction layer to perform feature extraction on the second concatenated feature to obtain a fourth fusion vector output by the last feature extraction layer in the at least one feature extraction layer; Weight the fourth fusion vector with the weights of each sample item in the multiple sample items respectively to obtain the target feature of each sample item; For any sample item, sequentially input the target feature of the sample item into the corresponding second multi-layer perceptron, self-attention layer, and adaptive layer to obtain the predicted click-through rate of the sample item output by the corresponding adaptive layer; Train the initial click-through rate model according to the predicted click-through rate and the true click-through rate of the sample item; 6. The method according to claim 5, characterized in that, The training the initial click-through rate model according to the predicted click-through rate and the true click-through rate of the sample item includes: Construct a maximum likelihood function according to the difference between the true click-through rate and the predicted click-through rate of the sample item; Solve the maximum likelihood function using maximum likelihood estimation to obtain target parameters; Use the target parameters as the model parameters of the trained click-through rate model; 7. The method according to claim 1, characterized in that, The feature processing layer includes an FM sub-model and a DCN sub-model. The using the feature processing layer in the initial click-through rate model to perform first feature processing on the first fusion vector to obtain a second fusion vector includes: Use the FM sub-model to perform dimensionality reduction processing on the first fusion vector to obtain a first feature, obtain multiple low-order features of the first feature, and fuse the multiple low-order features to obtain a first sub-fusion feature; Use the DCN sub-model to perform dimensionality increase processing on the first fusion vector to obtain a second feature, obtain multiple high-order features of the second feature, and fuse the multiple high-order features to obtain a second sub-fusion feature; Fuse the first sub-fusion feature and the second sub-fusion feature to obtain the second fusion feature; 8. An item recommendation method, characterized in that, Applied to the server side, it includes: Receive a target request sent by the client, and according to the second object feature of the target object in the target request, obtain a second collaborative feature and a second item feature sequence associated with the target object, where the second collaborative feature is obtained by performing behavior statistics on target items associated with the historical behavior of the target object, and the second item feature sequence includes item features of multiple target items; Screen out a second material for displaying the multiple target items that matches the historical behavior of the target object from the set item material library; According to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material, predict the click-through rates of the multiple target items through the trained click-through rate model to obtain the predicted click-through rates of the multiple target items; wherein, the click-through rate model is trained according to the method described in any one of claims 1 to 7; Determine the recommendation scores of the multiple target items according to the predicted click-through rates of the multiple target items; Send the multiple target items to the client according to the recommendation scores of the multiple target items, where the target sequence is used for the client to display the multiple target items according to the recommendation scores of the multiple target items.

9. The method according to claim 8, wherein Predict the click-through rates of the multiple target items through a trained click-through rate model according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material to obtain the predicted click-through rates of the multiple target items, including: Obtain a second historical behavior sequence within a second set historical period associated with the target object; Input the second object feature, the second collaborative feature, the second item feature sequence, the second material feature sequence, and the second historical behavior sequence into the trained click-through rate model to obtain the predicted click-through rates of the multiple target items output by the trained click-through rate model.

10. A training device for a click-through rate model, characterized in that, Including: An acquisition module, configured to acquire a first object feature, the first collaborative feature, and a first item feature sequence of a sample object, where the first collaborative feature is obtained by performing behavior statistics on sample items associated with the historical behavior of the sample object, and the first item feature sequence includes item features of multiple sample items associated with the sample object; A first fusion module, configured to fuse the first user feature, the first collaborative feature, and the first item feature sequence by using a vector fusion layer in an initial click-through rate model to obtain a first fusion vector; A first processing module, configured to perform first feature processing on the first fusion vector by using a feature processing layer in the initial click-through rate model to obtain a second fusion vector; A second processing module, configured to determine the weights of the multiple sample items according to the first fusion vector, the second fusion vector, and the first material feature sequence of the first material for displaying the multiple sample items; A training module, configured to train the initial click-through rate model according to the weights of the multiple sample items.

11. An item recommendation device, characterized in that, Applied to a server, including: A receiving module, configured to receive a target request sent by a client, and acquire a second collaborative feature and a second item feature sequence associated with the target object according to the second object feature of the target object in the target request, where the second collaborative feature is obtained by performing behavior statistics on target items associated with the historical behavior of the target object, and the second item feature sequence includes item features of multiple target items; A screening module, configured to screen out a second material for displaying the multiple target items that matches the historical behavior of the target object from a set item material library; A prediction module, configured to predict the click-through rates of the multiple target items according to the second object feature, the second collaborative feature, the second item feature sequence, and the second material feature sequence of the second material by using a click-through rate model trained by the method according to any one of claims 1 to 7 to obtain the predicted click-through rates of the multiple target items; A determination module, configured to determine a recommendation score of the multiple target items according to the predicted click-through rates of the multiple target items; A sending module, configured to send the multiple target items to a client according to the recommendation scores of the multiple target items, where the target sequence is used for the client to display the multiple target items according to the recommendation scores of the multiple target items.

12. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the training method of the click-through rate model according to any one of claims 1-7, or implements the item recommendation method according to any one of claims 8-9.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the click-through rate model according to any one of claims 1-7, or implements the item recommendation method according to any one of claims 8-9.

Images