A Click-Through Rate Prediction Method Based on Multi-Task Joint Learning

Abstract

The present invention belongs to the field of e-commerce big data recommendation, and particularly relates to a click-through rate prediction method based on multi-task joint learning, including obtaining product information and constructing user-product interest; according to the obtained product information, obtaining a product node sequence and a user node sequence, and calculating the second-order similarity between nodes in each respective sequence and their adjacent nodes; projecting the product nodes and user nodes onto the same feature space, calculating the attention coefficient of a node in this space, and obtaining the aggregated representation of the node using this coefficient; using the user-product interest to perform aggregated weighting on the nodes to obtain the final embedded representation of the nodes, embedding the product information and user information of the final embedded representation into a predictor, and predicting the probability of a user purchasing a product through the predictor; the present invention can assist e-commerce platforms in mastering the distribution of the behavioral characteristics of consumer groups, which is beneficial for advertisers to accurately place advertising information.

Description

Technical Field

[0001] The present invention belongs to the field of e-commerce big data recommendation, and particularly relates to a click-through conversion rate prediction method based on multi-task joint learning. Background Art

[0002] In recent years, the rapid development of the Internet has brought more and more convenience to people's lives. Among them, the recommendation system is an indispensable part of Internet applications. The recommendation system is applied in e-commerce platforms to enable users to obtain items they are interested in in a timely manner. At the same time, in the past few years, a new industry, online advertising, has emerged. There are two essential tasks in building a recommendation system: click-through rate (CTR) prediction and post-click conversion rate (CVR) prediction. Among them, CTR represents the conversion probability of a product from appearance to being clicked, and CVR represents the probability that a user will purchase a product after it is clicked.

[0003] On the other hand, with the development of e-commerce, online advertising has become one of the effective methods for merchants to promote products and obtain income. According to reports, in 2020, the total global online advertising expenditure reached 303.3 billion US dollars, accounting for 54.4% of the global advertising industry expenditure. For advertisers, in addition to highlighting the brand image, the more important thing in creating advertisements is to pursue the actual benefits of each advertisement placement, that is, whether users will purchase the products shown in the advertisement after the advertisement is displayed. Therefore, one of the key points of display advertising targeted placement is to target the advertisement to those exposure opportunities (Ad Impression) that will generate benefits as much as possible. This requires the advertisement placement system to predict the CVR before placement and, according to the prediction results, target the advertisement to the exposure opportunities with a high conversion rate to maximize the benefits. However, it is very difficult to accurately predict CVR because the user click conversion behavior occurs less frequently than the click-through behavior, and the conversion behavior may have a longer delay relative to the click behavior, which makes offline modeling very difficult. However, in the industry, by analyzing and researching CVR, not only can the market benefits after advertisement placement be judged, but also the impact of advertisement placement on the behavior of online consumers can be further clarified, helping advertisers better carry out marketing activities and achieve business goals. Therefore, CVR prediction has important commercial value and academic research value and has become an important field in recent years in both academia and industry.

[0004] With the continuous increase in commodity data and user behavior data on e-commerce platforms and the continuous development of graph neural networks, heterogeneous graph neural networks that can learn the relationships between different entities have been applied to the behavior prediction models on e-commerce platforms. Wang et al. (Wang X, Ji H, Shi C, et al. Heterogeneous graph attention network[C] / / The world wide web conference. 2019:2022-2032.) proposed a graph neural network that can simultaneously consider node-level attention and semantic-level attention. This network represents nodes under the same meta-path with weights, simultaneously considers the importance of nodes on each meta-path, and finally aggregates the two to obtain the final node representation. Inspired by this paper, the present invention proposes a click-through rate prediction method for user behavior on e-commerce platforms, introducing rich auxiliary information and historical behavior information of users in e-commerce platforms, which can not only effectively predict the click-through rate of commodities, but also analyze the impact of different user behaviors on the final click-through rate.

[0005] Based on the current research situation of click-through rate on e-commerce platforms, it is found that there are still some challenges in click-through rate prediction:

[0006] 1. Under e-commerce platforms, user effective behavior data is sparse. Although the amount of data generated by users under e-commerce platforms is very objective, the behavior data that can reflect the degree of users' preference for commodities is sparse. Mining effective user behaviors can improve the accuracy of click-through rate prediction tasks.

[0007] 2. Users have multiple behaviors towards a certain commodity. For commodities of interest, users may perform multiple behaviors such as clicking, adding to cart, collecting, and commenting on them. Each behavior represents the degree of users' preference for the commodity. Discussing the impact of different user behaviors on whether to place an order finally is also an important factor in improving prediction accuracy.

[0008] 3. Under e-commerce platforms, whether a user is willing to buy a commodity is independent of whether the commodity is clicked. However, when performing traditional model training, to judge whether a commodity can be finally purchased, the set of clicked commodities is used as the training space instead of the set of browsed commodities. This leads to sample selection bias, which results in a certain deviation in the final click-through rate. How to solve the sample selection bias problem is an important difficulty in studying click-through rate. Summary of the Invention

[0009] In response to the above challenges, the present application proposes a click-through rate prediction method based on multi-task joint learning, including the following steps:

[0010] Obtain user information, product information, auxiliary information, and user historical behavior data information under an e-commerce platform. The auxiliary information includes the user's search terms, store information, and event information that are directly associated with the user or the product.

[0011] Construct a user-product interest degree based on the user historical behavior data information.

[0012] Based on the auxiliary information, obtain a user node sequence and a product node sequence through the meta-path random walk strategy, that is, define a meta-path with nodes 1 - auxiliary information 1 - node 2 - auxiliary information 2 - … - node n - 1 - auxiliary information n - 1 - node n, and remove the auxiliary information in the meta-path to obtain the node sequence, where the node is a product node or a user node.

[0013] Calculate the second-order similarity between the nodes in the obtained user node sequence and product node sequence and their adjacent nodes.

[0014] Project the product nodes and user nodes onto the same feature space, specify a meta-path composed of nodes in this feature space, calculate the importance of one node to another node in this path, and normalize the importance of all neighbor nodes of the current node to the current node to obtain the attention coefficient of the current node.

[0015] Calculate the attention coefficient of the current node under a set of meta-paths, and use this coefficient to weighted sum the representations of the neighbor nodes of the current node in the projected feature space as the aggregated representation of the current node.

[0016] Use the user-product interest degree to weight the aggregated representations of the nodes in all original paths to obtain the final embedding representation of the node.

[0017] Obtain the user's rating matrix for products, and input the information of users and products obtained from the user click matrix and user purchase matrix through this matrix into a predictor to calculate the click-through rate and conversion rate of users for products.

[0018] Predict the probability that a user will purchase a certain product based on the click-through rate and conversion rate.

[0019] Furthermore, the construction of the user-product interest degree from the user historical behavior data information is expressed as:

[0020]

[0021] Among them, represents user u m for product i nDegree of interest; IsBuy, IsCollect, IsCart, IsClick, and IsBrowse respectively represent whether the user purchases, collects, adds to the shopping cart, clicks, or browses the product; when the user purchases the product, the value of IsBuy is 5; when the user collects the product, the value of IsCollect is 4; when the user adds the product to the shopping cart, the value of IsCart is 3; when the user clicks on the product, the value of IsClick is 2; when the user browses the product, the value of IsBrowse is 1; when the user does not purchase, collect, add to the shopping cart, click, or browse the product, the values of IsBuy, IsCollect, IsCart, IsClick, and IsBrowse are all 0.

[0022] Further, in the meta-path random walk strategy, given the heterogeneous network information and specifying the meta-path as the walk constraint, a corresponding node sequence is generated through the random walk strategy based on the meta-path. The probability of the random walk is expressed as:

[0023]

[0024] where represents the probability of the random walk, v m+1 represents the next node to be generated, represents a node of type A t , Φ s is the s-th specified meta-path; represents among the surrounding neighbors of the node the set of nodes of type A t+1 nodes, || represents finding the number of elements in the set; E represents the edges in the heterogeneous graph; φ(v m+1 ) is used to determine the type of the next node; is the set of nodes in the heterogeneous network, and ε is the set of edge relationships between nodes in the heterogeneous network.

[0025] Further, the final embedding representation of the node is expressed as:

[0026]

[0027] where Z = Z u ∪Z i , represents the final embedding representation of the node, represents the final embedding representation of user u m , represents the final embedding representation of product i q ; represents the s-th meta-path Φ sThe node embedding representation in; Denote at the sth meta-path Φ s The user u in m For the commodity i n The degree of interest; at the sth meta-path Φ s The magnitude of the user's interest in the commodity; S is the number of meta-paths.

[0028] Furthermore, the attention-weighted representation of a node i under the meta-path Φ s Is:

[0029]

[0030] Where Denote performing K times of learning, Is the representation of a node i under the meta-path Φ s K is the number of repeated trainings; σ() is the activation function; Denote the neighborhood of the node i under the meta-path Φ s ; Denote in the meta-path Φ s The attention coefficient of each neighbor node of the node i to it; h′ j Denote the vector representation h of a node j After being projected to the feature space where the commodity and the user are consistent, h j Is the commodity node or the user node.

[0031] Furthermore, the attention coefficient of each neighbor node of the node i to it is expressed as:

[0032]

[0033]

[0034] Where, [h′ i ||h′ j Denote concatenating the vector representation h′ of the node i i With the vector representation h′ of the node j j , h′ i And h′ j Are the vector representations h of the node i And the vector representation h of the node j After being projected to the same feature space, the node is the commodity node or the user node; Is a conversion coefficient, making [h′ i ||h′ j Convert from a matrix to a scalar; att node Denote the meta-path Φ sA deep neural network that performs node attention aggregation operations under Sim ij Represents the second-order similarity between node i and node j.

[0035] Furthermore, when calculating the second-order similarity between node i and node j, the calculation is performed based on whether the node is a user node or a product node. If the node is a user node, the user-user second-order similarity is expressed as:

[0036]

[0037] If the node is a product node, the second-order similarity between products is expressed as:

[0038]

[0039] in, For user u m With user u n The second-order similarity of |T u_mn | represents user u m and user u n The auxiliary information set T between u_mn The amount of auxiliary information in ; Indicates that the kth auxiliary information appears in user u m and user u n Auxiliary information set T u_mn The number of times in Indicates the kth auxiliary relationship in user u m and user u n The proportion of auxiliary relationship; For product i p and product i q The second-order similarity of |T i_pq |Indicates product i p and product i q Auxiliary information set T i_pq The amount of auxiliary information in ; Indicates that product i appears in the kth auxiliary information p and product i q Auxiliary information set T i_pq The number of times in Indicates the kth auxiliary relationship in product i p and product i q The proportion in the auxiliary relationship.

[0040] Furthermore, building a prediction model includes:

[0041]

[0042] in, Denote user u m The predicted probability of purchasing product i q ; Denote user u m The vector representation of Denote product i q The vector representation of Denote the final embedding representation of user u m ; Denote product i q The final embedding representation of and Denote product i q and user u m The implicit factors of; α and β are adjustment factors.

[0043] Furthermore, the loss function adopted during the training of the prediction model is expressed as:

[0044]

[0045] where £ is the loss function; Denote the set of meta-paths; u denotes the set of users; i denotes the set of products; Denote the vector representation of user u m ; Denote user u m The actual probability set of purchasing product i q ; λ is the first regularization parameter.

[0046] Furthermore, SGD is adopted to optimize the loss function, expressed as:

[0047]

[0048] where is The optimized value; is The optimized value; η is the learning rate.

[0049] The present invention predicts the probability of whether a user places an order, which can be applied to the recommendation system of an e-commerce platform, assisting the e-commerce platform to master the distribution of the behavior characteristics of the consumer group, and at the same time facilitating the accurate placement of advertising information by advertisers and reducing advertising costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic flowchart of a click-through rate prediction method based on multi-task joint learning according to the present invention;

[0051] Figure 2Schematic diagram of the heterogeneous graph attention network structure adopted by a click-through rate prediction method based on multi-task joint learning according to the present invention;

[0052] Figure 3 Schematic diagram of multi-task joint learning adopted in a click-through rate prediction method based on multi-task joint learning according to the present invention;

[0053] Figure 4 Schematic diagram of the prediction process of a click-through rate prediction method based on multi-task joint learning according to the present invention. Detailed implementation manners

[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0055] The present invention provides a click-through rate prediction method based on multi-task joint learning, which specifically includes the following steps:

[0056] Obtain user information, product information, auxiliary information, and user historical behavior data information under the e-commerce platform. The auxiliary information includes the search terms, store information, and activity information of the user directly associated with the user or product;

[0057] Construct the user-product interest degree through the user historical behavior data information;

[0058] Based on the auxiliary information, obtain the user node sequence and the product node sequence through the meta-path random walk strategy, that is, define a meta-path of node 1 - auxiliary information 1 - node 2 - auxiliary information 2 -... - node n - 1 - auxiliary information n - 1 - node n, and remove the auxiliary information in the meta-path to obtain the node sequence, where the node is a product node or a user node;

[0059] Calculate the second-order similarity between the nodes in each respective sequence and their adjacent nodes according to the obtained user node sequence and product node sequence;

[0060] Project the product nodes and user nodes onto the same feature space, specify a meta-path composed of nodes in this feature space, calculate the importance of one node to another node in this path, and normalize the importance of all neighbor nodes of the current node to the current node to obtain the attention coefficient of the current node;

[0061] Calculate the attention coefficient of the current node under a set of meta-paths, and use this coefficient to weighted sum the representations of the neighbor nodes of the current node in the projected feature space as the aggregated representation of the current node;

[0062] Use the user-item interest degree to weight the aggregated representations of the nodes in all original paths to obtain the final embedding representation of the node;

[0063] Obtain the rating matrix of users for items, and input the information of users and items obtained through the click matrix and purchase matrix of users into the predictor respectively to calculate the click-through rate and conversion rate of users for items;

[0064] Predict the probability that a user will purchase a certain item based on the click-through rate and conversion rate.

[0065] In this embodiment, the solution is mainly described from three aspects: data acquisition, relevant attribute extraction, and model establishment.

[0066] In the process of obtaining online data, the data source can be directly queried in real time from the database provided by the enterprise. It includes the basic feature data of users, the basic feature data of items, the historical behavior data of users (such as browsing, clicking, purchasing, adding to cart, collecting, etc.), and auxiliary information such as store information and marketing activity information under the e-commerce platform. The original data can be obtained through the public API of the social network or directly download the existing data source. Since the original data usually obtained is unstructured and cannot be directly used for data analysis, most unstructured data can be structured through simple data cleaning, for example, cleaning illegal or invalid information, deleting duplicate data, etc.

[0067] In the process of extracting relevant attributes, relevant attributes are extracted from four aspects: user basic feature data, item basic feature data, auxiliary information, and user behavior data. Under the e-commerce platform, whether a user finally places an order is affected by many factors, such as: the user's interests and hobbies, the effective behaviors of similar users, and the interaction behaviors of the user with the auxiliary information under the e-commerce platform, such as search terms, participation in marketing activities, etc. Based on this, the factors affecting user behavior are formalized in this paper. Different behaviors of users represent different degrees of preference for items by users. Therefore, the user-item interest degree is defined in this paper as:

[0068]

[0069] Among them, represents user u m for item i nDegree of interest; IsBuy, IsCollect, IsCart, IsClick, and IsBrowse respectively represent whether the user purchases, collects, adds to the shopping cart, clicks, or browses the product. When the user purchases the product, the value of IsBuy is 5; when the user collects the product, the value of IsCollect is 4; when the user adds the product to the shopping cart, the value of IsCart is 3; when the user clicks on the product, the value of IsClick is 2; when the user browses the product, the value of IsBrowse is 1; when the user does not purchase, collect, add to the shopping cart, click, or browse the product, the values of IsBuy, IsCollect, IsCart, IsClick, and IsBrowse are all 0.

[0070] In an e-commerce platform, users and products can often be associated through information such as brand, style, product category, etc. This article refers to this information as auxiliary information. Auxiliary information can be various association information between two nodes. For example, if two users participate in the same shopping activities (such as Double Eleven, 618, etc.), it can be considered that there is an association between the two users. Auxiliary information can also be that two users collect the same store, etc. In this embodiment, all information that connects two nodes is collectively referred to as auxiliary information. And second-order similarity can be established between user-user or product-product through auxiliary information. By mining the second-order similarity of nodes, the generalization ability of the model can be greatly improved, and the impact brought by the sparsity of user behavior data can be reduced. Therefore, this article defines the node second-order similarity as:

[0071]

[0072] In the formula, T represents the set of auxiliary relationships between two nodes, represents the influence magnitude of the auxiliary information k on the node i, and T k represents the number of times the auxiliary relationship k (or the k-th auxiliary relationship) appears between the node i and the node j; represents whether the node i and the node j jointly have a connection with the auxiliary information k. If so, the value is 1, otherwise it is 0.

[0073] The model building process is mainly divided into three stages. The first stage is to represent the UI node sequence based on auxiliary information. This paper extracts auxiliary information that can connect users and products from basic information data, and generates a more closely connected UI node sequence by designing a meta-path random walk strategy; the second stage is node information aggregation. This paper uses a heterogeneous graph neural attention network to calculate the importance of neighbor nodes to the current node, and granulates the importance of each node in the behavioral meta-path. Finally, the nodes are aggregated at the node level and semantic level; the third stage is to establish a multi-task joint learning MT-MF click-through conversion rate prediction model, and use matrix decomposition to construct a rating predictor in each auxiliary task to predict whether the user will complete the order behavior.

[0074] In the first stage, the method of walking representation guided by meta-path can solve the complex problem of multi-type nodes or multi-type edges in heterogeneous graphs. This paper designs a new random walk strategy based on meta-path to accurately represent node information. And specify the meta-path Φ as the walk constraint, and generate the corresponding node sequence through the random walk strategy based on the meta-path. The probability formula of random walk is shown as follows:

[0075]

[0076] in, Indicates A t Type of node, Indicates that the node Among the neighbors around A t+1 The number of type nodes.

[0077] Considering the diversity of user behavior, the same node may have different vector representations in different scenarios. Therefore, different edge types are assigned to the same node according to different scenarios. At the same time, auxiliary information from the e-commerce platform is used to find similarities between nodes in different scenarios.

[0078] Since this paper focuses on the prediction study of the click-through rate of users on products, it mainly represents user nodes and product nodes. Therefore, when designing the random walk paths, given the length of the node sequence, only user nodes or product nodes are used as the starting nodes of the meta-path. After generating the node sequence, the node sequence is traversed, and nodes of different types from the starting nodes are removed. In this way, it can be ensured that each obtained node sequence contains only either user nodes or product nodes. The advantage of doing this is that not only stronger associated user-user and product-product node sequences are obtained under auxiliary information, but also the original complex heterogeneous information network is transformed into a relatively simple homogeneous network, greatly reducing the complexity of the model.

[0079] After obtaining the node sequence, to explore the influence of different auxiliary information on nodes under the node sequence, the second-order similarity representations of user-user and product-product are obtained respectively. The second-order similarity of user-user is expressed as:

[0080]

[0081] The second-order similarity of product-product is expressed as:

[0082]

[0083] Among them, is the second-order similarity between user u m and user u n ; |T u_mn | represents the number of auxiliary information in the auxiliary information set T m between user u n and user u u_mn ; represents the number of times the k-th auxiliary information appears in the auxiliary information set T m between user u n and user u u_mn ; represents the proportion of the k-th auxiliary relationship in the auxiliary relationship between user u m and user u n ; is the second-order similarity between product i p and product i q ; |T i_pq | represents the number of auxiliary information in the auxiliary information set T p between product i q and product i i_pq ; represents the number of times the k-th auxiliary information appears in the auxiliary information set T p between product i q and product i i_pq ; Indicates the proportion of the k-th auxiliary relationship in the auxiliary relationships of product i p and product i q in the auxiliary relationship.

[0084] Determining the influence of different auxiliary information on nodes according to the magnitude of the second-order similarity of nodes completes the further representation of the node sequence.

[0085] As Figure 2 shown, in order to more realistically express the influence of different user behaviors on the final click-through conversion rate, we introduce a Heterogeneous Graph Attention Network (HAN for short) to learn node-level attention and semantic-level attention. In Figure 2 , the user-product interest degree is used to weight the aggregated representation of nodes on all original paths to obtain the final embedding representation of the node.

[0086] Since there is heterogeneity between user nodes and product nodes, and different types of nodes have different feature spaces, we need to set a transformation matrix for user nodes and product nodes to project the features of these two types of nodes onto the same feature space. The conversion process is expressed as:

[0087] h′ i = M · h i

[0088] where h i and h′ i are the original features and projected features of node i respectively, and M is the transformation matrix. After type conversion, node-level attention can handle nodes of both user and product types simultaneously.

[0089] In the heterogeneous graph attention network, given two nodes v i , v j under the specified meta-path Φ, the importance of node v j to node v i can be learned through the self-attention mechanism as shown in the following formula:

[0090]

[0091] where h′ i , h′ j are the vector representations of v i , v j respectively; att node represents the deep neural network that performs node attention aggregation operations under this meta-path, and att nodeIt is applicable to the node pairs under the entire meta-path, so the importance of any node to each node under the meta-path can be obtained. The central node v in the commodity node sequence generated in the above section i For example, define its neighbor nodes as in Indicates that v under the meta-path Φ i The neighbor set (including v i ). Then, find v i Each neighbor node v j v i Importance And normalize it, as shown in the following formula, to get each neighbor node pair v i Attention coefficient Expressed as:

[0092]

[0093] Among them, σ represents the activation function, || represents the concatenation of the features of two nodes, is a unit vector, which is used to make the concatenated [h′ i ||h′ j ] is converted into a scalar, and those skilled in the art can convert [h′] into a scalar according to the specific implementation process. i ||h′ j ] dimension settings The dimension of this application is not limited to this; Sim ij Indicates the second-order similarity between node i and node j. The value of this parameter is calculated based on the node type (product node, user node). The calculation method for user nodes is as follows: Calculation reference for product nodes Then we get the neighbor node to node v under this meta-path. i The multi-head attention coefficient makes the model more stable. A node i in the meta-path Φ s The attention weighted expression under is:

[0094]

[0095] in, Indicates K times of learning, is the meta-path Φ s The representation of a node i under , K is the number of repeated training; σ() is the activation function; Represents the meta-path Φ s The neighborhood of the next i-node; Indicates that in the meta-path Φ s The attention coefficient of each neighbor node of node i.

[0096] Similarly, for a set of meta-paths we can obtain the corresponding node multi-head attention representation vectors So far, we have completed the attention aggregation at the node level.

[0097] Because the relationships between users and products are nothing more than operations such as browsing, clicking, and purchasing. Therefore, according to the user-product interest degree UCH, the attention coefficients at the semantic level of the meta-path are represented. Finally, the node v i The final embedding representation is expressed as:

[0098]

[0099] where, Z = Z u ∪Z i , representing the final embedding representation of the node, representing the final embedding representation of user u m ; representing the final embedding representation of product i q ; representing the node embedding representation in the s-th meta-path Φ s ; representing the degree of interest of user u s in product i m in the s-th meta-path Φ n ; the magnitude of the degree of interest of the user in the product in the s-th meta-path Φ s ; S is the number of meta-paths.

[0100] To avoid the inconsistency between the training space and the inference space, resulting in sample selection bias and affecting the final result as Figure 3 shown, we adopt the method of multi-task joint learning to obtain the final click-through rate PCvr. As Figure 4 , in this embodiment, auxiliary tasks of click-through rate PCtr and conversion rate PCtcvr are established to improve the accuracy of the main task PCvr.

[0101] In the auxiliary tasks, the classic matrix factorization (MF) is used to obtain the ratings of users for products to predict PCtr and PCtcvr. In MF, the rating of user u for product i can be simply defined as follows:

[0102]

[0103] where, represents the predicted probability that user u m purchases product i q ; represents the vector representation of user u m ; Denote commodity i q as the vector representation. Incorporate the final embedded representations Z u of the user and the commodity and Z i into the rating predictor, then the predicted probability m that user u q purchases commodity i can be expressed as:

[0104]

[0105] wherein, and represent the implicit factors of commodity i q and user u m ; α and β are adjustment factors.

[0106] The vector representation m of user u and the vector representation q of commodity i take different values when calculating the click-through rate PCtr and the conversion rate PCtcvr. During the process of calculating the click-through rate PCtr, obtain the rating matrix of the user for the commodity. In this rating matrix, it is the rating of the user for the commodity, and the score is between 0 and 5 (including 0 and 10). A score of 2 to 4 indicates that the user has clicked on the commodity. Different scores are assigned according to the type of click (click types such as adding to the shopping cart, favoriting, etc.). For example, in this embodiment, let the score for the user to browse be 1, the score for clicking be 2, the score for adding to the shopping cart be 3, the score for favoriting be 4, and the score for purchasing be 5. When the user clicks, that is, when performing operations such as adding to the shopping cart, favoriting, and purchasing, a click behavior on the commodity also occurs. That is, when constructing the click matrix of the user, the score of this part of the user is between 2 and 5, then set the value of this part of the user to 1, and the value of other users to 0, convert it into a click matrix, and input the information of the user and the commodity in this matrix into the predictor to calculate the click-through rate PCtr; if the score of the user is 5, that is, the user has clicked and purchased, then set the value of this part of the user to 1, and the value of other users to 0, convert it into a click-purchase matrix, and input the information of the user and the commodity in this matrix into the predictor to calculate the conversion rate PCtcvr. For example, in Figure 4 , the obtained user rating matrix is expressed as:

[0107]

[0108] wherein,? represents an unknown rating; the above user rating matrix is converted into a click matrix, that is, set the ratings with values of 2 to 5 to 1, and the others to 0, and the unknown ratings remain unchanged, which is expressed as:

[0109]

[0110] The above user rating matrix is converted into a click matrix, that is, the ratings with a value of 5 are set to 1, others are set to 0, and the unknown ratings remain unchanged, which is expressed as:

[0111]

[0112] The user information and product information in the click matrix and the click-to-purchase matrix are input into the predictor to calculate the click-through rate PCtr and the conversion rate PCtcvr. In this embodiment, the calculated click-through rate PCtr is 0.6 and the conversion rate PCtcvr is 0.2. Finally, the probability that a user purchases a certain product is predicted based on the click-through rate and the conversion rate, that is, the click conversion rate PCvr, which is expressed as:

[0113]

[0114] During the training of the predictor, that is, during the training of the prediction model, the loss function is used for backpropagation to optimize the parameters of the predictor. The loss function is expressed as:

[0115]

[0116] Among them, £ is the loss function; represents the set of meta-paths; u represents the set of users; i represents the set of products; represents the vector representation of user u m ; represents the actual probability set that user u m purchases product i q ; λ is the first regularization parameter.

[0117] To prevent the phenomenon of overfitting, SGD is used in this embodiment to optimize the final objective function. The optimization process is expressed as:

[0118]

[0119] Among them, is the optimized value; is the optimized value; η is the learning rate.

[0120] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A click-through rate prediction method based on multi-task joint learning, characterized in that The steps include: Obtain user information, product information, auxiliary information, and user historical behavior data information under an e-commerce platform. The auxiliary information includes the search terms of users, store information, and activity information that are directly associated with the users or products. Construct user-product interest degree through user historical behavior data information. Based on the auxiliary information, obtain user node sequences and product node sequences through the meta-path random walk strategy, that is, define a meta-path of node 1 - auxiliary information 1 - node 2 - auxiliary information 2 - … - node n - 1 - auxiliary information n - 1 - node n, and remove the auxiliary information in the meta-path to obtain the node sequence, where the nodes are product nodes or user nodes. Project the product nodes and user nodes onto the same feature space, specify a meta-path composed of nodes in this feature space, calculate the importance of one node to another node in this path, and normalize the importance of all neighbor nodes of the current node to the current node to obtain the attention coefficient of the current node. Calculate the attention coefficient of the current node under a group of meta-paths, and use this coefficient to weighted sum the representations of the neighbor nodes of the current node in the projected feature space as the aggregated representation of the current node. Use the user-product interest degree to weight the aggregated representations of the nodes in all original paths to obtain the final embedded representation of the node. Obtain the user rating matrix for products, and input the information of users and products obtained through this matrix to obtain the user click matrix and user purchase matrix into a predictor to calculate the click-through rate and conversion rate of users for products. Predict the probability that a user will purchase a certain product through the click-through rate and conversion rate.

2. The click-through rate prediction method based on multi-task joint learning according to claim 1, characterized in that The construction of user-product interest degree from user historical behavior data information is expressed as: Among them, represents the degree of interest of user u m in product i n ; IsBuy, IsCollect, IsCart, IsClick, and IsBrowse respectively represent whether the user purchases, collects, adds to the shopping cart, clicks, or browses the product; when the user purchases the product, the value of IsBuy is 5; when the user collects the product, the value of IsCollect is 4; when the user adds the product to the shopping cart, the value of IsCart is 3; when the user clicks on the product, the value of IsClick is 2; when the user browses the product, the value of IsBrowse is 1; when the user does not purchase, collect, add to the shopping cart, click, or browse the product, the values of IsBuy, IsCollect, IsCart, IsClick, and IsBrowse are all 0.

3. A click-through rate prediction method based on multi-task joint learning according to claim 1, characterized in that In the meta-path random walk strategy, given the heterogeneous network information and specifying the meta-path as the walk constraint, the corresponding node sequence is generated through the random walk strategy based on the meta-path. The probability of the random walk is expressed as: Among them, represents the probability of random walk, v m+1 represents the next node to be generated, represents a node of type A t , Φ s is the s-th meta-path formulated; represents among the surrounding neighbors of node the set of nodes of type A t+1 , || represents the number of elements in the set; E represents the edges in the heterogeneous graph; φ(v m+1 ) represents judging the type of the next node, that is, finding the type of node v m+1 ; is the set of nodes in the heterogeneous network, and ε is the set of edge relationships between nodes in the heterogeneous network.

4. A click-through rate prediction method based on multi-task joint learning according to claim 1, characterized in that, The final embedded representation of the node is expressed as: Where Z = Z u ∪Z i , represents the final embedding representation of the node, Represents user u m The final embedding representation of Indicates product i q The final embedding representation of represents the s-th element path Φ s Node embedding representation in ; Indicates that the s-th element path Φ s User u m For product i n The interest level of the s-th element path Φ s The user's interest in the product; S is the number of meta-paths.

5. A click-through rate prediction method based on multi-task joint learning according to claim 4, characterized in that, The attention-weighted representation of a node $i$ under the meta-path $\varPhi$ s is given by: Among them, represents performing K times of learning, is the representation of a node i under the meta-path Φ s , where K is the number of repeated training; σ() is the activation function; represents the neighborhood of node i under the meta-path Φ s ; represents the attention coefficient of each neighbor node of node i to it in the meta-path Φ s ; h′ j represents the vector representation h of a node j projected onto the vector representation after the feature space where the commodity and the user are consistent, h j is a commodity node or a user node.

6. A click-through rate prediction method based on multi-task joint learning according to claim 5, characterized in that The attention coefficient of each neighbor node of node i to it is expressed as: Among them, [h′ i ||h′ j represents concatenating the vector representation h′ i of node i with the vector representation h′ j of node j. h′ i and h′ j are the vector representations h i of the node and h j of the node after projecting them onto the same feature space. The node is a commodity node or a user node; is a conversion coefficient that converts [h′ i ||h′ j from a matrix to a scalar; att node represents the deep neural network that performs node attention aggregation operation under the meta-path Φ s ; Sim ij represents the second-order similarity between node i and node j.

7. A click-through rate prediction method based on multi-task joint learning according to claim 6, characterized in that, When calculating the second-order similarity between node i and node j, calculate according to whether the node is a user node or a product node. If the node is a user node, the user-user second-order similarity is expressed as: If the node is a product node, the product-product second-order similarity is expressed as: Among them, is the second - order similarity between user u m and user u n ; |T u_mn | represents the number of auxiliary information in the auxiliary information set T m between user u n and user u u_mn ; represents the number of times the k - th auxiliary information appears in the auxiliary information set T m between user u n and user u u_mn ; represents the proportion of the k - th auxiliary relationship in the auxiliary relationships between user u m and user u n ; is the second - order similarity between product i p and product i q ; |T i_pq | represents the number of auxiliary information in the auxiliary information set T p between product i q and product i i_pq ; represents the number of times the k - th auxiliary information appears in the auxiliary information set T p between product i q and product i i_pq ; represents the proportion of the k - th auxiliary relationship in the auxiliary relationships between product i p and product i q .

8. A click-through rate prediction method based on multi-task joint learning according to claim 1, characterized in that The construction of the prediction model includes: Among them, represents the predicted probability that user u m buys commodity i q ; represents the vector representation of user u m ; represents the vector representation of commodity i q ; represents the final embedding representation of user u m ; represents the final embedding representation of commodity i q ; and represent the implicit factors of commodity i q and user u m ; α and β are adjustment factors.

9. A click-through rate prediction method based on multi-task joint learning according to claim 8, characterized in that The loss function used in the process of training the prediction model is expressed as: Among them, £ is the loss function; represents the meta-path set; u represents the user set; i represents the commodity set; represents the vector representation of user u m ; represents the actual probability set that user u m purchases commodity i q ; λ is the first regularization parameter.

10. A click-through rate prediction method based on multi-task joint learning according to claim 9, characterized in that, Use SGD to optimize the loss function, which is expressed as: Among them, is the optimized value; is the optimized value; η is the learning rate.

Images