A method for predicting product click-through rate by fusing product features across platforms

Abstract

This invention pertains to personalized recommendation technology for e-commerce platforms, specifically a method for predicting product click-through rates (CTR) through cross-platform product feature fusion. The method includes: extracting product feature information from the target platform to guide product migration from the source platform; obtaining different vectors based on user behavior and relationship information, shared product sets, and private product sets, and embedding and concatenating them into a single feature vector; segmenting user groups based on user attributes and the similarity of products with potential interests to construct similar user groups; calculating the independent CTR prediction for each user within a group, and aggregating them to obtain the user's overall predicted CTR for that product. This invention, by segmenting the product set of the source platform and utilizing a knowledge graph of products and behaviors, embeds and concatenates different types of vectors to obtain a more complete final vector, enabling a more accurate product recommendation sequence.

Description

Technical Field

[0001] This invention belongs to the field of personalized recommendation technology for e-commerce platforms, specifically relating to a method for predicting product click-through rates by integrating cross-platform product features. Background Technology

[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 lives and making network services more popular, widespread, and diversified. Among the many network services, e-commerce stands out, not only providing people with a convenient shopping experience but also helping users quickly find products that match their interests and needs through personalized recommendation technology. Personalized recommendations not only make it easier for users to find products they like but also bring more business opportunities and value to e-commerce platforms. Mining user behavior data based on long-term and short-term interests to achieve personalized recommendations has received widespread attention from scholars both domestically and internationally.

[0003] To date, researchers both domestically and internationally have conducted extensive research on product recommendation. Existing product recommendation methods can be broadly categorized into two main types: recommendation methods based on user attributes and historical behavior data, and recommendation methods based on complex models using machine learning and deep learning. The former measures the similarity between users by leveraging publicly available attribute information on social platforms and the topological structure of social networks; the latter extracts user features through feature engineering, trains a classification model based on labeled data, and finally uses the trained model to predict personalized recommendations.

[0004] In summary, numerous researchers have made significant achievements in personalized research, laying a solid foundation for further studies. However, research on cross-platform user recommendations using marketing data from actual e-commerce platforms still faces several challenges:

[0005] Problem 1: The data distribution across platforms is inconsistent, with varying feature distributions and data scales, leading to unsatisfactory cross-platform feature fusion results. Different e-commerce platforms emphasize different product categories, and users tend to buy different types of products on different platforms, resulting in an imbalanced sample distribution. A holistic understanding of user interests and preferences is key to more realistically simulating user behavior. Only when the model understands user behavior can it better predict future user actions.

[0006] Question 2: Significant differences in user shopping preferences across platforms severely impact the comprehensive representation of user interests and preferences. E-commerce platform service strategies guide user consumption, leading to variations in user preferences across different platforms. User data from a single e-commerce platform is insufficient to fully represent user interest characteristics, resulting in biased click-through rate predictions.

[0007] Question 3: E-commerce platforms suffer from sparse effective data, making it difficult to obtain. Existing user and product interaction data is sparse relative to the overall dataset, causing the model to fail to achieve ideal results and thus impacting the prediction model. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for predicting product click-through rates by fusing product features across platforms, comprising:

[0009] S1: Extract product feature information from the target platform to guide the migration of products from the source platform. Divide the product set of the source platform 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: Based on user behavior and relationship information, shared product set and private product set, different vectors are obtained. Then, multiple vectors are embedded and concatenated to finally obtain a feature vector containing user information, shared preferences and private preferences.

[0011] S3: Segment users, build similar user groups, enhance the information on the characteristics of products of potential user interest, and reduce the complexity of matching users and products;

[0012] S4: After reducing the complexity of matching users and products, perform independent click-through rate prediction calculations for each user in the group based on feature vectors containing user information, shared preferences, and private preferences, and then aggregate them to obtain the comprehensive click-through rate prediction value for the user for the product.

[0013] The beneficial effects of this invention are:

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

[0015] This invention achieves more granular cross-platform feature fusion by assigning different mapping matrices to different operational products, making the recommendation results closer to the real situation. At the same time, by fusing cross-platform user historical behavior sequences, it can achieve more accurate cross-platform product click-through rate prediction. Attached Figure Description

[0016] Figure 1 This is a flowchart of a product click-through rate prediction method that realizes cross-platform product feature fusion according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] A method for predicting product click-through rates by fusing product features across platforms, such as... Figure 1 As shown, it includes:

[0019] S1: Extract product feature information from the target platform to guide the migration of products from the source platform. Divide the product set of the source platform 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: Based on user behavior and relationship information, shared product set and private product set, different vectors are obtained. Then, multiple vectors are embedded and concatenated to finally obtain a feature vector containing user information, shared preferences and private preferences.

[0021] S3: Segment users, build similar user groups, enhance the information on the characteristics of products of potential user interest, and reduce the complexity of matching users and products;

[0022] Similarity:

[0023]

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

[0025] S4: After reducing the complexity of matching users and products, perform independent click-through rate prediction calculations for each user in the group based on feature vectors containing user information, shared preferences, and private preferences, and then aggregate them to obtain the comprehensive click-through rate prediction value for the user for the product.

[0026] In this embodiment, the target platform's product feature information is extracted to guide the migration of products from the source platform. The source platform's product set is divided into a shared product set and a private product set, and different mapping matrices are assigned. Convolution and Hadamard product are combined to migrate different product sets from the source platform, including:

[0027] S11: Utilize user account information from e-commerce platforms to collect users' historical behavior information, including browsing, clicking, favorites, and purchase information. At the same time, collect information on users' friend relationships and transfer product features from the source platform to the feature space of the target platform.

[0028] S12: Divide the source platform's product set into a shared product set and a private product set G. shared and G private And assign independent mapping matrices.

[0029] The source platform's product collection is divided into shared product collections and private product collections, including:

[0030]

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

[0032] S13: Repeat steps S11 to S12 to migrate the user behavior information and product feature information from the source platform S to the feature space of the target platform, and then perform a full representation of the cross-platform feature information.

[0033] Combining target platform feature information C tgf Product feature information is transferred to the target platform feature space using two independent mapping matrices, including:

[0034]

[0035] Among them, X tgf X represents the feature matrix after the source platform's feature information has been transferred to the target platform. shared W represents the feature vector of shared goods. shared Represents the shared goods mapping matrix, X private W represents the feature vector of a privately owned commodity. private Represents a private goods mapping matrix. The symbol represents the Hadamard product, and * represents convolution.

[0036] In this embodiment, different vectors are obtained based on user behavior and relationship information, shared product sets, and private product sets. These vectors are then embedded and concatenated to ultimately obtain a feature vector containing user information, shared preferences, and private preferences, including:

[0037] By analyzing the historical user behavior sequences on each platform, and combining the platform to which the operated goods belong and the user, the source platform's product set is divided into a shared product set and a private product set. Each product set contains different types of user preference information, especially private preferences, which should be given more attention. In addition, the target platform may also contain private preference information. Product feature vectors are constructed from the shared product set and the private product set respectively.

[0038] User operations within each set of products are related in a certain way. The relatedness includes: similar products being operated on, high substitutability, and a product network being constructed by connecting related products. In the product network, nodes represent products, and edges represent the relationships between products.

[0039] By extracting user behavior sequences and social relationship features from across platforms, a user behavior sequence and relationship network are constructed. Similar users among different users can achieve feature complementarity, thereby constructing a user behavior embedding vector containing user-related information.

[0040] Finally, we constructed corresponding product knowledge graph views for the shared product set and the private product set, respectively. Vector embeddings were then used to embed user behavior. Embedding shared product embedding vectors on the basis of and private goods embedded vector Obtain the overall embedding vector containing user behavior information, shared product information, and private product information.

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

[0042] By constructing a product knowledge graph from shared and private product sets, including:

[0043] Within the product set, product attributes are extracted, with each product node including information such as brand, category, and price. Similar product nodes are connected, and a product knowledge graph is constructed using products as nodes and relationships between products as edges. Furthermore, based on users' different historical behavior sequences and social relationship information, a user knowledge graph is constructed using users as nodes and user behaviors and relationships between users as edges.

[0044] The final vector is constructed by embedding different categories of user behavior into three views: shared product view, private product view, and user behavior view. For each view layer, the strategy of aggregating neighbor node information in HAN is used to embed the vector. The final result is a holistic embedding vector containing user behavior information, shared product information, and private product information.

[0045] Construct shared product embedding vectors for both the shared product set and the private product set. and private goods embedded vector include:

[0046]

[0047]

[0048] Among them, S i P represents a shared collection of goods that interact with each other. i Represents a private collection of goods for interaction. This indicates the interaction weight between the user and the shared goods. e represents the interaction weight between the user and the private goods. j e represents the embedding vector of shared goods. k This represents the embedding vector of a private product.

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

[0050]

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

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

[0053]

[0054] in, Represents the behavior matrix in the network. Represents user behavior embedding vectors, This indicates the user's historical behavior sequence information; This represents the shared or private set of goods information operated under the current action, and σ() represents the multi-view feature vector embedding function.

[0055] In this embodiment, user groups are segmented based on user attributes and the similarity of potential interest items to construct similar user groups. This enhances the feature information of potential interest items and reduces the complexity of matching users and items, including:

[0056]

[0057] Among them, Group i Indicates the final user group group number for user i; argmin j This indicates selecting the group that minimizes the average distance; d n (User i Cluster j ) is the distance metric between user i and group j in the nth dimension; N is the total number of feature dimensions, covering user similarity-related feature information such as basic user attributes, behavioral information, social relationships, and product feature information.

[0058] Based on feature vectors containing user information, shared preferences, and private preferences, independent click-through rate (CTR) predictions are calculated for each user within the group. These CTRs are then aggregated to obtain a comprehensive CTR prediction value for the user's viewpoint on the product, including:

[0059] Unique click-through rate:

[0060]

[0061] Overall click-through rate:

[0062]

[0063] Among them, X U Group represents a feature vector containing user information, shared preferences, and private preferences. i Indicates the user's user group, W U This represents the weight matrix corresponding to the user, where b is the offset and w is the weight matrix. i Let y represent the weight of the i-th user. U This represents user U's predicted value for the product. This represents the predicted overall click-through rate of the product, where N represents the total number of feature dimensions, and T represents the matrix transpose.

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

Claims

1. A method for predicting product click-through rate by fusing product features across platforms, characterized in that, include: S1: Extract product feature information from the target platform to guide product migration from the source platform; Extracting product feature information from the target platform to guide product migration from the source platform, including: S11: Extracting target platform feature information using convolutional properties C tgf It is used to guide the migration of product feature information from the source platform; S12: Divide the source platform's product set into a shared product set and a private product set, and combine this with the target platform's characteristic information C. tgf Product feature information is transferred to the target platform feature space using two independent mapping matrices. The source platform's product collection is divided into shared product collections and private product collections, including: Wherein, D(o) i (S, T) represents whether a product in the source platform belongs to a shared product set or a private product set, where a value of 1 indicates a shared product set and a value of 0 indicates a private product set; S and T represent the source platform and the target platform, respectively; i Indicates the goods that need to be judged; Combining target platform feature information C tgf Product feature information is transferred to the target platform feature space using two independent mapping matrices, including: Among them, X tgf X represents the feature matrix after the source platform's feature information has been transferred to the target platform. shared W represents the feature vector of shared goods. shared Represents the shared goods mapping matrix, X private W represents the feature vector of a privately owned commodity. private represents the private commodity mapping matrix, ° represents the Hadamard product, and * represents convolution; S13: Migrate user behavior information and user social relationship information from the source platform to the feature space of the target platform; S2: Based on user behavior and relationship information, shared product set and private product set, different vectors are obtained. Multiple vectors are embedded and concatenated to obtain a feature vector containing user information, shared preferences and private preferences. S3: Segment users based on their own attributes and the similarity of products of potential interest, build similar user groups, enhance the feature information of products of potential interest, and reduce the complexity of matching users and products; S4: After reducing the complexity of matching users and products, perform independent click-through rate prediction calculations for each user in the group based on feature vectors containing user information, shared preferences, and private preferences, and then aggregate them to obtain the comprehensive click-through rate prediction value for the user for the product.

2. The method for predicting product click-through rate by fusing product features across platforms 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 sequences of users on each platform, and combining the platform to which the operated goods belong and the operating user, the source platform's product set is divided into a shared product set and a private product set. Each product set contains different types of user preference information. User operations within each set of products are related, and the relatedness includes: similar products and high substitutability. A product network is constructed by connecting related products, in which nodes represent products and edges represent the relationships between products. In a product network, shared product embedding vectors are constructed for both shared product sets and private product sets. and private goods embedded vector By extracting user behavior sequences and social relationship features from across platforms, a user behavior sequence and relationship network are constructed. Similar users among different users can achieve feature complementarity, thereby constructing a user behavior embedding vector containing user-related information. Construct corresponding product knowledge graph views for shared and private product sets respectively, and embed vector u into user behavior. i(friends) Embedding shared goods vector u on the basis i(shared) and private goods embedded vector u i(private) This yields an overall embedding vector containing user behavior information, shared product information, and private product information.

3. The method for predicting product click-through rate by fusing product features across platforms according to claim 2, characterized in that, Construct shared product embedding vectors for both the shared product set and the private product set. and private goods embedded vector include: Among them, S i P represents a shared collection of goods that interact with each other. i Represents a private collection of goods for interaction. This indicates the interaction weight between the user and the shared goods. e represents the interaction weight between the user and the private goods. j e represents the embedding vector of shared goods. k This represents the embedding vector of a private product.

4. The product click-through rate prediction method for cross-platform product feature fusion as described in claim 2, characterized in that, Construct user behavior embedding vectors containing user-related information include: Among them, B i B j Let w represent the behavioral feature sequences of users i and l respectively. il F represents the similarity weight between friends. i Let α represent the set of friends, and let α represent the weight of the behavioral features.

5. The method for predicting product click-through rate by fusing product features across platforms according to claim 2, characterized in that, A product knowledge graph is constructed by integrating shared and private product sets, and vector data is embedded into user behavior. Embedding shared product embedding vectors on the basis of and private goods embedded vector Obtain the overall embedding vector containing user behavior information, shared product information, and private product information. include: Product attributes are extracted from the product set. Each product node includes brand, category and price information. Similar product nodes are connected, and a product knowledge graph is constructed with products as nodes and relationships between products as edges. In addition, based on users’ different historical behavior sequence information and social relationship information, a user knowledge graph is constructed with users as nodes and user behavior and relationships between users as edges. Constructing the final vector: For the three views—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 view. Embedded representations are performed for different categories of user behavior to obtain an overall embedding vector containing user behavior information, shared product information, and private product information.

6. The method for predicting product click-through rate by fusing product features across platforms according to claim 5, characterized in that, For the three views—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 view. Embedded representations are performed for different categories of user behavior to obtain an overall embedding vector containing user behavior information, shared product information, and private product information. include: in, Represents the behavior matrix in the network. Represents user behavior embedding vectors, This indicates the user's historical behavior sequence information; This represents the shared or private set of goods information operated under the current action, and σ() represents the multi-view feature vector embedding function.

7. The method for predicting product click-through rate by fusing product features across platforms according to claim 1, characterized in that, Based on user attributes and the similarity of potential interest products, user groups are segmented to construct similar user groups, enhancing the feature information of potential interest products and reducing the complexity of matching users and products, including: Among them, Group i Indicates the final user group group number for user i; argmin j This indicates selecting the group that minimizes the average distance; d n (User i Cluster j ) represents the distance metric between user i and group j in the nth dimension; N represents the total number of feature dimensions.

8. The method for predicting product click-through rate by fusing product features across platforms according to claim 1, characterized in that, Based on feature vectors containing user information, shared preferences, and private preferences, independent click-through rate (CTR) predictions are calculated for each user within the group. These CTRs are then aggregated to obtain a comprehensive CTR prediction value for the user's viewpoint on the product, including: Unique click-through rate: Overall click-through rate: Among them, X U Group represents a feature vector containing user information, shared preferences, and private preferences. i Indicates the user's user group, W U This represents the weight matrix corresponding to the user, where b is the offset and w is the weight matrix. i Let y represent the weight of the i-th user. U This represents user U's predicted value for the product. This represents the predicted overall click-through rate of the product, where N represents the total number of feature dimensions, and T represents the matrix transpose.

Images