Commodity click-through rate prediction method for realizing cross-platform commodity feature fusion

Abstract

The invention belongs to an e-commerce platform personalized recommendation technology, and particularly relates to a commodity click rate prediction method for realizing cross-platform commodity feature fusion, which comprises the following steps: extracting commodity feature information of a target platform to guide commodity migration of a source platform; obtaining different vectors according to the user behaviors, the relation information, the shared commodity set and the private commodity set, and performing embedding and splicing to form a feature vector; dividing the user groups according to the attributes of the users and the similarity of the potential interested commodities, and constructing similar user groups; and carrying out independent click rate prediction calculation on each user in the group, and carrying out aggregation to obtain a comprehensive click rate prediction value of the user for the commodity. According to the method, the commodity set of the source platform is divided, and different types of vectors are embedded and spliced to obtain the final vector with more complete information through the knowledge graph of the commodities and behaviors, so that a more accurate commodity recommendation sequence can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of personalized recommendation on e-commerce platforms, and specifically relates to a method for predicting click-through rate of commodities by realizing cross-platform commodity feature fusion. Background Art

[0002] With the rapid development of Internet technology and the widespread application of mobile smart devices, various online social media, e-commerce platforms, video entertainment and other network service platforms have emerged rapidly, greatly enriching people's online life, making network services more popular, extensive and diversified. Among the many network services, the e-commerce field is particularly prominent. It not only provides people with a convenient shopping experience, but also helps users quickly find products that meet their interests and needs through personalized recommendation technology. Personalized recommendation not only provides convenience for users to find their favorite products, but also brings more business opportunities and value to e-commerce platforms. Mining user behavior data based on long-term and short-term interests of users to achieve personalized recommendations has received widespread attention from scholars at home and abroad.

[0003] So far, domestic and foreign researchers have conducted a lot of research on product recommendation. Existing product recommendation methods can be generally divided into two categories: recommendation methods based on user attributes and historical behavior data and recommendation methods based on machine learning and deep learning complex models. The former measures the similarity between users by using the attribute information disclosed on the social platform and the topological structure of the social network; the latter extracts user features through feature engineering, trains the classification model based on the labeled data, and finally uses the trained model to predict personalized recommendations.

[0004] In summary, many researchers have made many achievements in personalization research, laying a solid foundation for further research. There are still some challenges in the process of cross-platform user recommendation based on the marketing data of actual e-commerce platforms:

[0005] Problem 1: There are differences in data distribution across platforms, and the feature distribution and data scale are not uniform, resulting in unsatisfactory cross-platform feature fusion effects. Different e-commerce platforms focus on different types of goods, and users tend to buy different types of goods on different platforms, resulting in unbalanced sample distribution. Globally grasping user interest preferences is the key to more realistically simulating user behavior. Only by allowing the model to understand user behavior can we make better predictions about the user's next behavior;

[0006] Problem 2: The differences in user shopping preferences across platforms seriously affect the overall representation of user interest preferences. The e-commerce platform service strategy guides user consumption, resulting in differences in user preferences on different platforms. The user data of a single e-commerce platform is not sufficient to fully represent the user's interest characteristics, resulting in deviations in click-through rate predictions;

[0007] Problem 3: The effective data of e-commerce platforms is sparse and difficult to obtain. The existing user and product interaction data is sparse compared to the overall data set, which makes the model unable to achieve the ideal effect, thus affecting the prediction model. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides a method for predicting the click-through rate of a product by realizing cross-platform product feature fusion, comprising:

[0009] S1: Extract the target platform product feature information to guide the source platform product migration, divide the source platform product set into a shared product set and a private product set, assign different mapping matrices, and combine convolution and Hadamard product to migrate different product sets of the source platform;

[0010] S2: Different vectors are obtained based on user behavior and relationship information, shared product sets, and private product sets. Multiple vectors are then embedded and concatenated to finally obtain a feature vector containing user information, shared preferences, and private preferences.

[0011] S3: Divide user groups, build similar user groups, enhance the characteristic information of products that users are potentially interested in, and reduce the complexity of matching between users and products;

[0012] S4: After reducing the complexity of matching between users and products, an independent click rate prediction calculation is performed for each user in the group based on a feature vector containing user information, shared preferences, and private preferences, and then aggregated to obtain the user's comprehensive click rate prediction value for the product.

[0013] Beneficial effects of the present invention:

[0014] The present invention solves the problem of unsatisfactory recommendation results caused by partial data loss after cross-platform and cross-domain user behavior sequence fusion, and provides a comprehensive cross-platform representation of user's potential interest products and behaviors, thereby reducing the feature dimensions of users and products and the complexity of data processing.

[0015] The present invention assigns different mapping matrices to different operation products, realizes cross-platform feature fusion in a finer granularity, and makes the recommendation results closer to the actual situation. At the same time, through the fusion of cross-platform user historical behavior sequences, more accurate cross-platform product click-through rate prediction can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of a method for predicting the click-through rate of a product by integrating product features across platforms. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] A method for predicting product click-through rate by integrating product features across platforms, such as Figure 1 As shown, including:

[0019] S1: Extract the target platform product feature information to guide the source platform product migration, divide the source platform product set into a shared product set and a private product set, assign different mapping matrices, and combine convolution and Hadamard product to migrate different product sets of the source platform;

[0020] S2: Different vectors are obtained based on user behavior and relationship information, shared product sets, and private product sets. Multiple vectors are then embedded and concatenated to finally obtain a feature vector containing user information, shared preferences, and private preferences.

[0021] S3: Divide user groups, build similar user groups, enhance the characteristic information of products that users are potentially interested in, and reduce the complexity of matching between users and products;

[0022] Similarity:

[0023]

[0024] Among them, U and P represent two feature vectors whose similarity needs to be calculated, corresponding to the similarity of interest products between different users. If the similarity is greater than u, they are classified into one user group.

[0025] S4: After reducing the complexity of matching between users and products, an independent click rate prediction calculation is performed for each user in the group based on a feature vector containing user information, shared preferences, and private preferences, and then aggregated to obtain the user's comprehensive click rate prediction value for the product.

[0026] In this embodiment, the target platform commodity feature information is extracted to guide the source platform commodity migration, the source platform commodity set is divided into a shared commodity set and a private commodity set, different mapping matrices are allocated, and convolution and Hadamard product are combined to migrate different commodity sets of the source platform, including:

[0027] S11: Using the user account information of the e-commerce platform, collect the user's historical behavior information, including browsing, clicking, collecting and purchasing operation information of goods, and collect the user's friend relationship information, and migrate the source platform product features to the target platform feature space;

[0028] S12: Divide the source platform product set into a shared product set and a private product set G shared and G private , and allocate independent mapping matrices.

[0029] Divide the source platform product collection into a shared product collection and a private product collection, including:

[0030]

[0031] Among them, D(o i , S, T) indicates whether a product in the source platform belongs to the shared product set or the private product set, where the value 1 indicates the shared product set and the value 0 indicates the private product set; S and T represent the source platform and the target platform respectively; o i Indicates the product that needs to be judged;

[0032] S13: Repeat steps S11 to S12 to migrate the user behavior information and product feature information on the source platform S to the feature space of the target platform, and then present the cross-platform feature information in its entirety;

[0033] Combined with the target platform feature information C tgf The product feature information is migrated to the target platform feature space through two independent mapping matrices, including:

[0034]

[0035] Among them, X tgf represents the feature matrix after the feature information of the source platform is migrated to the target platform, X shared The feature vector representing the shared goods, W shared represents the shared goods mapping matrix, X private The feature vector representing the private goods, W private represents the private product mapping matrix, represents the Hadamard product and * represents convolution.

[0036] In this embodiment, different vectors are obtained according to user behavior and relationship information, shared product sets and private product sets, and then multiple vectors are embedded and spliced ​​to finally obtain a feature vector containing user information, shared preferences and private preferences, including:

[0037] By analyzing the historical operation behavior sequence of users on each platform, the source platform product set is divided into a shared product set and a private product set in combination with the platform to which the operated product belongs and the operated user. Each product set contains different types of user preference information, especially private preferences should increase more interest. In addition, the target platform may also have private preference information. The product feature vectors are constructed through the shared product set and the private product set.

[0038] The user operations in each divided commodity set have certain correlations, which include: the operated commodities are similar and highly substitutable, and a commodity network is constructed by connecting the related commodities. In the commodity network, nodes represent commodities and edges represent the correlation between commodities.

[0039] By extracting cross-platform user behavior sequences and social relationship feature information, constructing user behavior sequences and relationship networks, similar users between different users can achieve feature complementarity, thereby constructing a user behavior embedding vector containing user-related information.

[0040] Finally, we construct corresponding product knowledge graph views for the shared product set and the private product set. Embed shared goods embedding vector and private item embedding vector Get the overall embedding vector containing user behavior information, shared product information and private product information

[0041] In this embodiment, a commodity network is constructed by connecting related commodities, including:

[0042] The product knowledge graph is constructed by combining shared product sets and private product sets, including:

[0043] Extract product attributes from the product set. Each product node includes information such as brand, category, and price. Connect similar product nodes and build a product knowledge graph with products as nodes and relationships between products as edges. In addition, based on different historical behavior sequence information and social relationship information of users, build a user knowledge graph with users as nodes and user behaviors and relationships between users as edges.

[0044] Construct the final vector. For the three views of shared product view, private product view and user behavior view, the strategy of aggregating neighborhood node information in HAN is used in each layer of view to embed different categories of user behaviors. Finally, the overall embedding vector containing user behavior information, shared product information and private product information is obtained.

[0045] Construct shared product embedding vectors for shared product sets and private product sets respectively and private item embedding vector include:

[0046]

[0047]

[0048] Among them, S i represents the set of shared products for interaction, P i represents the collection of private goods for interaction, represents the interaction weight between users and shared products, represents the interaction weight between the user and the private product, e j represents the embedding vector of the shared item, e k Embedding vector representing private items.

[0049] Construct a user behavior embedding vector containing user-related information include:

[0050]

[0051] Among them, B i , B j Respectively represent the behavioral feature sequences of users i and j, and w ij represents the similarity weight between friends, F i represents the friend set, and α represents the behavior feature weight.

[0052] In this embodiment, a shared feature vector and a private feature vector are embedded in the user behavior vector to obtain a final vector, including:

[0053]

[0054] in, represents the behavior matrix in the network, represents the user behavior embedding vector, Indicates the historical behavior sequence information of the user; represents the shared product set or private product set information operated under the current behavior, and σ() represents the multi-view feature vector embedding function.

[0055] In this embodiment, based on the user's own attributes and the similarity of potential interest products, the user groups are divided, similar user groups are constructed, the characteristic information of the user's potential interest products is enhanced, and the complexity of matching between users and products is reduced, including:

[0056]

[0057] Among them, Group i Indicates the final user group group number of user i; argmin j Indicates the selection of the group with the minimum average distance; d n (User i , Cluster j ) is the distance measure between user i and group j in the nth dimension; N is the total number of feature dimensions, covering feature information related to user similarity such as user basic attributes, behavior information, social relationships, and product feature information.

[0058] Based on the feature vector containing user information, shared preferences and private preferences, an independent click rate prediction calculation is performed for each user in the group, and the user's comprehensive click rate prediction value for the product is obtained by aggregation, including:

[0059] Unique click rate:

[0060]

[0061] Comprehensive click rate:

[0062]

[0063] Among them, X U Represents a feature vector containing user information, shared preferences, and private preferences. i Indicates the user group to which the user belongs, W U represents the weight matrix corresponding to the user, b is the offset, w i represents the weight of the i-th user, y U represents the predicted value of user U for the predicted product, It represents the predicted value of the comprehensive click-through rate of the predicted product, N represents the total number of feature dimensions, and T represents matrix transpose.

[0064] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting product click-through rate by realizing cross-platform product feature fusion, characterized in that: include: S1: Extract the target platform's commodity feature information to guide the source platform's commodity migration; S2: Different vectors are obtained based on user behavior and relationship information, shared product sets, and private product sets. Multiple vectors are embedded and concatenated to obtain a feature vector containing user information, shared preferences, and private preferences. S3: Divide user groups according to user attributes and similarity of potential interest items, build similar user groups, enhance feature information of user potential interest items, and reduce the complexity of matching between users and items; S4: After reducing the complexity of matching between users and products, an independent click rate prediction calculation is performed for each user in the group based on a feature vector containing user information, shared preferences, and private preferences, and then aggregated to obtain the user's comprehensive click rate prediction value for the product.

2. A method for predicting click-through rate of commodities by realizing cross-platform commodity feature fusion according to claim 1, characterized in that: Extract the target platform's product feature information to guide the source platform's product migration, including: S11: Extract target platform feature information C using convolution features tgf , used to guide the migration of product feature information on the source platform; S12: Divide the source platform product set into a shared product set and a private product set, and combine the target platform feature information C tgf Migrate product feature information to the target platform feature space through two independent mapping matrices; Divide the source platform product collection into a shared product collection and a private product collection, including: Among them, D(o i ,S,T) indicates whether a product in the source platform belongs to the shared product set or the private product set, where the value 1 indicates the shared product set and the value 0 indicates the private product set; S and T represent the source platform and the target platform respectively; o i Indicates the product that needs to be judged; Combined with the target platform feature information C tgf The product feature information is migrated to the target platform feature space through two independent mapping matrices, including: X tgf =((X shared οC tgf )*W shared )+((X private οC tgf )*W private ) Among them, X tgf represents the feature matrix after the feature information of the source platform is migrated to the target platform, X shared The feature vector representing the shared goods, W shared represents the shared goods mapping matrix, X private The feature vector representing the private goods, W private represents the private commodity mapping matrix, ° represents the Hadamard product, and * represents the convolution; S13: Migrate the source platform user behavior information and user social relationship information to the target platform feature space.

3. The method for predicting the click-through rate of a product by realizing cross-platform product feature fusion according to claim 1, characterized in that: Different vectors are obtained based on user behavior and relationship information, shared product sets, and private product sets. Multiple vectors are embedded and concatenated to obtain a feature vector containing user information, shared preferences, and private preferences, including: By analyzing the historical operation behavior sequence of users on each platform, the source platform product set is divided into a shared product set and a private product set in combination with the platform to which the operated products belong and the operated users. Each product set contains different types of user preference information, especially private preferences should increase more interest. In addition, the target platform may also have private preference information. The user operations in each divided commodity set are all correlated, and the correlation includes: the operated commodities are similar and highly substitutable, and a commodity network is constructed by connecting the correlated commodities, in which nodes represent commodities and edges represent the correlation between commodities; In the commodity network, the shared commodity embedding vector u is constructed for the shared commodity set and the private commodity set respectively. i(shared) and the private product embedding vector u i(private) ; By extracting cross-platform user behavior sequences and social relationship feature information, we construct user behavior sequences and relationship networks. Similar users between different users can achieve feature complementarity, thereby constructing a user behavior embedding vector u containing user-related information. i(friends) ; The shared product set and the private product set are constructed into corresponding product knowledge graph views respectively, and the user behavior embedding vector u i(friends) Embed the shared product embedding vector u based on i(shared) and the private product embedding vector u i(private) , and obtain the overall embedding vector containing user behavior information, shared product information, and private product information 4. The method for predicting the click-through rate of a product by realizing cross-platform product feature fusion according to claim 3, characterized in that: Construct the shared product embedding vector u for the shared product set and the private product set respectively i(shared) and the private product embedding vector u i(private) , include: Among them, S i represents the set of shared products for interaction, P i represents the collection of private goods for interaction, represents the interaction weight between users and shared products, represents the interaction weight between the user and the private product, e j represents the embedding vector of the shared item, e k Embedding vector representing private items.

5. The method for predicting the click-through rate of a product by realizing cross-platform product feature fusion according to claim 3, characterized in that: Construct a user behavior embedding vector u containing user-related information i(friends) ,include: Among them, B i , B j Respectively represent the behavioral feature sequences of users i and j, and w ij represents the similarity weight between friends, F i represents the friend set, and α represents the behavior feature weight.

6. The method for predicting the click-through rate of a product by realizing cross-platform product feature fusion according to claim 3, characterized in that: By building a product knowledge graph for shared product sets and private product sets, and embedding the vector u in user behavior i(friends) Embed the shared product embedding vector u based on i(shared) and the private product embedding vector u i(private) , and obtain the overall embedding vector containing user behavior information, shared product information, and private product information include: Extract product attributes from the product set. Each product node includes brand, category and price information. Connect similar product nodes and build a product knowledge graph with products as nodes and relationships between products as edges. In addition, based on different historical behavior sequence information and social relationship information of users, build a user knowledge graph with users as nodes and user behaviors and relationships between users as edges. Construct the final vector: For the three views of shared product knowledge graph, private product knowledge graph and user knowledge graph, the strategy of aggregating neighborhood node information in HAN is adopted in each layer of view to embed different categories of user behaviors and obtain the overall embedding vector containing user behavior information, shared product information and private product information.

7. A method for predicting product click-through rate by realizing cross-platform product feature fusion according to claim 6, characterized in that: For the three views of shared product knowledge graph, private product knowledge graph and user knowledge graph, the strategy of aggregating neighborhood node information in HAN is adopted in each layer of view to embed different categories of user behaviors and obtain the overall embedding vector containing user behavior information, shared product information and private product information. include: in, Represents the behavior matrix in the network, u i(friends) represents the user behavior embedding vector, u i(behavior) Indicates the historical behavior sequence information of the user; u i(pri|sha) represents the shared product set or private product set information operated under the current behavior, and σ() represents the multi-view feature vector embedding function.

8. The method for predicting the click-through rate of a product by realizing cross-platform product feature fusion according to claim 1, characterized in that: According to the user's own attributes and the similarity of potential interest products, the user group is divided, similar user groups are constructed, the characteristic information of the user's potential interest products is enhanced, and the complexity of matching between users and products is reduced, including: Among them, Group i Indicates the final user group group number of user i; argmin j Indicates the selection of the group with the minimum average distance; d n (User i ,Cluster j ) represents the distance measure between user i and group j in the nth dimension; N represents the total number of feature dimensions.

9. The method for predicting the click-through rate of a product by realizing cross-platform product feature fusion according to claim 1, characterized in that: Based on the feature vector containing user information, shared preferences and private preferences, an independent click rate prediction calculation is performed for each user in the group, and the user's comprehensive click rate prediction value for the product is obtained by aggregation, including: Unique click rate: Comprehensive click rate: Among them, X U Represents a feature vector containing user information, shared preferences, and private preferences. Group i Indicates the user group to which the user belongs, W U represents the weight matrix corresponding to the user, b is the offset, w i represents the weight of the i-th user, y U represents the predicted value of user U for the predicted product, It represents the predicted value of the comprehensive click-through rate of the predicted product, N represents the total number of feature dimensions, and T represents matrix transpose.

Images