A social e-commerce recommendation method based on multi-view network purification
By employing a multi-view network cleansing method, a user-product interaction network and a user-user social network for a social e-commerce platform are constructed. Graph convolutional networks and adaptive denoising techniques are then used to solve the noise problem in social e-commerce recommendations, achieving more accurate personalized recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-10
AI Technical Summary
Existing social e-commerce recommendation methods are relatively simplistic in handling noisy information and struggle to adapt to complex real-world environments, resulting in unstable recommendation performance and impacting user experience and purchase rates.
A multi-view network cleansing method is adopted, which constructs a robust learning model of user and product representations through dual-view collaborative denoising, adaptive social network cleansing and cross-view comparative learning. This includes constructing a user-product interaction network and a user-user social network, extracting node representations using graph convolutional networks, filtering out false social edges by combining explicit-implicit denoising and adaptive denoising techniques, and aligning representations through comparative learning.
It effectively removes noise from social e-commerce platforms, improves recommendation accuracy and robustness, and significantly enhances user purchase experience and purchase rate.
Smart Images
Figure CN122367585A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of social recommendation technology, specifically a social e-commerce recommendation method based on multi-view network purification. Background Technology
[0002] With the rapid development of social e-commerce platforms, the user base and the number of products have continued to grow, and user behaviors such as browsing, collecting, purchasing, and sharing have become increasingly diverse. Recommendation systems have become one of the core technologies for improving the user shopping experience and increasing platform conversion rates. Social recommendations, by incorporating user social relationships and adhering to the theories of social homogeneity and social influence, effectively alleviate the data sparsity and cold start problems of traditional recommendation systems.
[0003] In recent years, graph neural networks (GNNs) have demonstrated powerful advantages in capturing graph-structured data representations. They aggregate neighborhood information to capture collaborative filtering signals and social influences related to user preferences, thereby achieving personalized recommendations. Therefore, social recommendation methods based on GNNs have become one of the mainstream technological directions. However, in real-world social e-commerce scenarios, user interactions with products often involve false interactions such as accidental clicks, temporary browsing, and fraudulent orders, deviating from genuine purchasing preferences. Furthermore, social networks contain false or unreliable social edges such as stranger followers, marketing accounts, and weakly connected friends, which can propagate erroneous preference signals. This noise is further amplified when using GNNs for neighborhood propagation, resulting in poor-quality learned user and product representations, thus impacting recommendation performance.
[0004] Existing social e-commerce recommendation methods generally employ simplistic noise handling approaches, often failing to adapt to complex real-world environments and maintain stable and reliable recommendation performance in noisy conditions. Therefore, constructing a recommendation method with robustness, generalization ability, and adaptability to social e-commerce scenarios, utilizing multi-view network cleansing, has become a crucial issue. Summary of the Invention
[0005] To overcome the shortcomings of existing social e-commerce recommendation platforms where noise interference leads to low recommendation accuracy, this invention proposes a social e-commerce recommendation method based on multi-view network purification. Through dual-view collaborative denoising, adaptive social network purification, and cross-view comparative learning, robust learning of user and product representations is achieved, thereby improving the accuracy of e-commerce recommendations.
[0006] The technical solution adopted by this invention to solve the technical problem is: A social e-commerce recommendation method based on multi-view network purification is proposed. First, a user-product interaction network and a user-user social network are constructed based on user behavior and social data from the social e-commerce platform. Then, a dual-view denoising generator is used to process the user-product interaction network using implicit denoising with a variational graph autoencoder and explicit denoising based on social enhancement, resulting in two complementary purified interaction networks. Next, adaptive denoising is applied to the social network, using semantic similarity and structural similarity as metrics to eliminate false social edges. Then, a dual-graph encoder is used to learn purified user and product representations, and aligned representations are learned through cross-view comparison. Finally, a recommendation score is calculated based on the fused representations to achieve personalized product recommendations.
[0007] Furthermore, the method includes the following steps: Step 1: Construct user-product interaction networks (UI networks) based on user behavior and social data on the social e-commerce platform. and users - user social networks ,in Represents a set of users. Represents a collection of goods. and Representing the number of users and products respectively, the adjacency matrix Adjacency matrix represents the set of interaction edges between users and products. This represents the set of social edges between users; Step 2: Initialize User Representation and product representation ,in As a representation dimension; Step 3: Extract node representations from the original UI network using a graph convolutional network.
[0008] Furthermore, the method also includes the following steps: Step 4: Calculate the posterior distribution parameters of the latent variables using two independent multilayer perceptrons. Step 5: Sample the latent representations of nodes from the posterior distribution;
[0009] Step 6: Reconstruct the interaction graph using an inner product decoder to obtain the purified interaction network. .
[0010] Furthermore, the method also includes the following steps: Step 7: Calculate the cosine similarity matrix between users, and then obtain the adjacency matrix corresponding to the purified social network; Step 8: Calculate the enhanced user representation; Step 9: Calculate the similarity matrix between users and products. This filters out low-similarity interaction edges to obtain a purified interaction network.
[0011] The method further includes the following steps: Step 10: Calculate the semantic similarity and structural similarity of all social edges in the social network; Step 11: Weighted fusion to obtain the overall credibility Then, a threshold filtering method is used to denoise the social network, resulting in a purified social network. .
[0012] The method further includes the following steps: Step 12: Use a parameter-sharing graph convolutional network to encode the cleaned UI network and social network respectively to obtain enhanced user and product representations; Step 13: Perform adaptive weighted representation fusion through an attention mechanism.
[0013] The method further includes the following steps: Step Fourteen: Calculate the contrastive learning loss ; Step 15: Aggregate the user and product representations from the two views to form the final representations of users and products.
[0014] Furthermore, the method also includes the following steps: Step 16: Calculate the generation loss in implicit denoising based on variational graph autoencoders. ; Step 17: Calculate the primary recommendation loss .
[0015] The method further includes the following steps: Step 18: Calculate the total loss:
[0016] in, , and For adjustable hyperparameters, This is the regularization loss; Step 19: Repeat steps 3 to 18 until the total loss is below the specified threshold, at which point the calculation ends and the final user and product representations are obtained. Step 20: Obtain a list of recommended products: For any target user on the social e-commerce platform Based on the representation it ultimately learns and all the product representations learned Calculate the user's choice of each candidate item Preference score This generates a list of recommended products for each user. .
[0017] The technical concept of this invention is as follows: This invention addresses the noise problem in the social network and UI network of social e-commerce platforms. It adopts explicit-implicit collaborative denoising to generate two complementary cleaned views, and integrates semantic and structural similarity as a metric to adaptively clean the social network. Then, through contrastive learning, it aligns multi-view representations to learn robust and general user and product representations, thereby improving the robustness and accuracy of recommendations.
[0018] The beneficial effects of this invention are: it can effectively remove the complex noise existing in social e-commerce platforms, effectively capture users' true purchasing intentions, and significantly improve users' own purchasing experience and purchase rate in social e-commerce platforms. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of a social e-commerce recommendation method based on multi-view network purification. Detailed Implementation
[0020] The present invention will now be further described with reference to the accompanying drawings.
[0021] Reference Figure 1 This paper presents a social e-commerce recommendation method based on multi-view network purification. This method effectively alleviates the widespread noise problem in social e-commerce platforms. It employs an explicit-implicit collaborative approach to purify the UI network and uses a fusion of semantic and structural similarity metrics for adaptive denoising of social networks. This enables the graph neural network to effectively learn more realistic user and product representations, thereby providing users with personalized product recommendations. The method includes the following steps: Step 1: Construct user-product interaction networks (UI networks) based on user behavior and social data on the social e-commerce platform. and users - user social networks ,in Represents a set of users. Represents a collection of goods. and Representing the number of users and products respectively, the adjacency matrix Adjacency matrix represents the set of interaction edges between users and products. This represents the set of social edges between users; In this embodiment, a user set is constructed by selecting 100 users with rich social connections from the social e-commerce platform. And 1000 items to build a product collection And build user-to-user social networks based on each user's social data. And the user-product interaction network is constructed through user interactions such as purchasing, adding to favorites, and recommending products. ; Step 2: Initialize User Representation and product representation ,in As a representation dimension; Step 3: Extract node representations from the original UI network using a graph convolutional network: ; in, For lightweight graph convolutional networks, For the original users and products The splicing of representations; Step 4: Calculate the posterior distribution parameters of the latent variables using two independent multilayer perceptrons: ; ; in, and These represent the mean and variance, respectively; Linear(.) is a linear transformation layer; and ReLU(.) and Softplus(.) are non-linear activation functions. Step 5: Sample the latent representations of nodes from the posterior distribution:
[0022] Step 6: Reconstruct the interaction graph using an inner product decoder, where... ; ; users respectively The reconstruction probabilities of corresponding positive and negative sample items, For inner product operations, For users The latent representation. and For users The corresponding positive and negative sample product latent representations, when Delete the interaction edge when it is less than a certain value. When the value exceeds a certain threshold, the interaction edge is filled in, thus purifying the interaction network. ; Step 7: Calculate the cosine similarity matrix between users: ; in, Assuming an initial user representation, we can then obtain the adjacency matrix corresponding to the purified social network: ; in, The preset hyperparameter threshold; Step 8: Calculate the enhanced user representation: ; in, To purify the corresponding angle matrix of social networks; Step 9: Calculate the similarity matrix between users and products: ; This filters out low-similarity interaction edges to obtain a purified interaction network. : ; in, For product characterization, The preset hyperparameter threshold, The adjacency matrix corresponds to the purified social network; Step 10: Calculate the semantic similarity and structural similarity of all social edges in the social network: ; ; in, Represents the set of neighbors of a node. and They represent and Semantic and structural similarity between them; Step 11: Weighted fusion to obtain the overall credibility : ; Next, a threshold filtering method is used. ; Denoising social networks results in a cleaner social network. ,in, It is an adjustable parameter. The preset hyperparameter threshold; Step 12: Using a parameter-sharing graph convolutional network, encode the purified UI network and social network respectively to obtain enhanced user and product representations: ; ; in, For the purified UI network ( (or social networks) D is the corresponding angle matrix, W is an optional trainable weight matrix, and Z is the enhanced user or product representation. Step 13: Adaptive weighted representation fusion using an attention mechanism: ; in, ; ; This indicates a splicing operation. It is a multilayer perceptron. To perform UI network noise reduction ( The user representations learned The user representation learned on the denoised social network; Step Fourteen: Calculate the contrastive learning loss: ; in ; ; in, (.) represents the cosine similarity function. and Users in explicit denoising view Fusion representation and commodity The representation, and For users in implicitly denoised views Fusion representation and commodity The representation, Temperature coefficient; Step 15: Aggregate the user and product representations from the two views to form the final user and product representations: ; ; Step 16: Calculate the generation loss in implicit denoising based on variational graph autoencoders: ; in ; ; in, and Let be the reconstruction probabilities between the nth user and its positive and negative sample items, respectively. To recommend losses, For KL divergence loss term, These are adjustable hyperparameters; Step 17: Calculate the primary recommendation loss: ; in For users Its positive sample products Similarity between them For users Its negative sample products Similarity between them Represents a nonlinear activation function; Step 18: Calculate the total loss: ; in, , and For adjustable hyperparameters, This is the regularization loss; Step 19: Repeat steps 3 to 18 until the total loss is below the specified threshold, at which point the calculation ends and the final user and product representations are obtained. Step 20: Obtain a list of recommended products: For any target user on the social e-commerce platform Based on the representation it ultimately learns and all the product representations learned Calculate the user's choice of each candidate item Preference score This generates a list of recommended products for each user. ; In this embodiment, based on the final calculated preference score between the user and the product, the top 20 products are recommended to the user.
[0023] In this embodiment, a social network is constructed by introducing social data from users on a social e-commerce platform and combined with a user-product interaction network. For the user-product interaction network, an explicit-implicit collaborative approach is used to denoise it, generating two complementary cleansing networks. For the social network, a fake social edge filtering method is applied, using semantic and structural similarity as metrics. Furthermore, a contrastive learning strategy is employed to align user and product representations with distributional shifts across different views. This method enables the social e-commerce platform not only to make recommendations based on users' historical interaction data but also to quickly generate personalized product recommendations for new users based on their social data, effectively alleviating the data sparsity and cold start problems.
[0024] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.
Claims
1. A social e-commerce recommendation method based on multi-view network purification, characterized in that, First, a user-product interaction network and a user-user social network are constructed based on user behavior and social data from a social e-commerce platform. Then, a dual-view denoising generator is used to process the user-product interaction network using implicit denoising with a variational graph autoencoder and explicit denoising based on social enhancement, resulting in two complementary clean interaction networks. Next, adaptive denoising is applied to the social network, using semantic similarity and structural similarity as metrics to eliminate false social edges. Then, a dual-graph encoder is used to learn clean user and product representations, and the representations are aligned through cross-view comparison. Finally, a recommendation score is calculated based on the fused representations to achieve personalized product recommendations.
2. The social e-commerce recommendation method based on multi-view network purification as described in claim 1, characterized in that, The method includes the following steps: Step 1: Construct user-product interaction networks (UI networks) based on user behavior and social data on the social e-commerce platform. and users - user social networks ,in Represents a set of users. Represents a collection of goods. and Representing the number of users and products respectively, the adjacency matrix Adjacency matrix represents the set of interaction edges between users and products. Represents the set of social edges between users; Step 2: Initialize User Representation and product representation ,in As a representation dimension; Step 3: Extract node representations from the original UI network using a graph convolutional network: ； in, For lightweight graph convolutional networks, For the original users and products The splicing of representations.
3. The social e-commerce recommendation method based on multi-view network purification as described in claim 1, characterized in that, The method further includes the following steps: Step 4: Calculate the posterior distribution parameters of the latent variables using two independent multilayer perceptrons: ； ； in, and These represent the mean and variance, respectively; Linear(.) is a linear transformation layer; and ReLU(.) and Softplus(.) are non-linear activation functions. Step 5: Sample the latent representations of nodes from the posterior distribution: Step 6: Reconstruct the interaction graph using an inner product decoder, where... ； ； users respectively The reconstruction probabilities of corresponding positive and negative sample items, For inner product operations, For users The latent representation. and For users The corresponding positive and negative sample product latent representations, when Delete the interaction edge when it is less than the set lower threshold. When the value exceeds a set upper threshold, the interaction edge is filled in, thus purifying the interaction network. .
4. The social e-commerce recommendation method based on multi-view network purification as described in claim 3, characterized in that, The method further includes the following steps: Step 7: Calculate the cosine similarity matrix between users: ； in, Assuming an initial user representation, we can then obtain the adjacency matrix corresponding to the purified social network: ； in, The preset hyperparameter threshold; Step 8: Calculate the enhanced user representation: ； in, To purify the corresponding angle matrix of social networks; Step 9: Calculate the similarity matrix between users and products: ； This filters out low-similarity interaction edges to obtain a purified interaction network. : ； in, For product characterization, The preset hyperparameter threshold, This is the adjacency matrix corresponding to the purified social network.
5. The social e-commerce recommendation method based on multi-view network purification as described in claim 4, characterized in that, The method further includes the following steps: Step 10: Calculate the semantic similarity and structural similarity of all social edges in the social network: ； ； in, Represents the set of neighbors of a node. and They represent and Semantic and structural similarity between them; Step 11: Weighted fusion to obtain the overall credibility : ； Next, a threshold filtering method is used. ； Denoising social networks results in a cleaner social network. ,in, It is an adjustable parameter. This is the preset hyperparameter threshold.
6. The social e-commerce recommendation method based on multi-view network purification as described in claim 5, characterized in that, The method further includes the following steps: Step 12: Using a parameter-sharing graph convolutional network, encode the purified UI network and social network respectively to obtain enhanced user and product representations: ； ； in, For the purified UI network ( (or social networks) D is the corresponding angle matrix, W is an optional trainable weight matrix, and Z is the enhanced user or product representation. Step 13: Adaptive weighted representation fusion using an attention mechanism: ； in, ； ； This indicates a splicing operation. It is a multilayer perceptron. The user representation learned on the denoised UI network. This refers to the user representation learned on a denoised social network.
7. The social e-commerce recommendation method based on multi-view network purification as described in claim 6, characterized in that, The method further includes the following steps: Step Fourteen: Calculate the contrastive learning loss: ； in ； ； in, (.) represents the cosine similarity function. and Users in explicit denoising view Fusion representation and commodity The representation, and For users in implicitly denoised views Fusion representation and commodity The representation, Temperature coefficient; Step 15: Aggregate the user and product representations from the two views to form the final user and product representations: ； 。 8. The social e-commerce recommendation method based on multi-view network purification as described in claim 7, characterized in that, The method further includes the following steps: Step 16: Calculate the generation loss in implicit denoising based on variational graph autoencoders: ； in ； ； in, and Let be the reconstruction probabilities between the nth user and its positive and negative sample items, respectively. To recommend losses, For KL divergence loss term, These are adjustable hyperparameters; Step 17: Calculate the primary recommendation loss: ； in For users Its positive sample products Similarity between them For users Its negative sample products Similarity between them This represents a non-linear activation function.
9. The social e-commerce recommendation method based on multi-view network purification as described in claim 8, characterized in that, The method further includes the following steps: Step 18: Calculate the total loss: ； in, , and For adjustable hyperparameters, This is the regularization loss; Step 19: Repeat steps 3 to 18 until the total loss is below the specified threshold, at which point the calculation ends and the final user and product representations are obtained. Step 20: Obtain a list of recommended products: For any target user on the social e-commerce platform Based on the representation it ultimately learns and all the product representations learned Calculate the user's choice of each candidate item Preference score This generates a list of recommended products for each user. 。