Social recommendation method and system based on denoising and preference balance

Abstract

The invention discloses a social recommendation method and system based on denoising and preference balance, and belongs to the technical field of graph neural networks and social recommendation. The method comprises the following steps: constructing a social graph; denoising the social graph to obtain a denoised social graph; taking the normalized user item interaction matrix and the de-noised social graph as input, performing double-branch variational coding learning probabilistic user embedding, and performing variational regularization, gating de-noising and attention fusion to obtain user fusion representation; based on user fusion representation, joint reconstruction of an interaction graph and a social graph is carried out, and a training target is obtained through a preference balance optimization target. According to the method, through adaptive graph structure optimization and a long-tail preference constraint mechanism, accurate modeling of a real social relation and balanced expression of diversified preferences are realized, and finally, the reliability, fairness and recommendation precision of the model in a noise sparse social environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph neural networks and social recommendation, and particularly relates to a social recommendation method and system based on denoising and preference balancing. BACKGROUND

[0002] GraphRec is a graph neural network model for social recommendation tasks, and its core idea is to model the user-item interaction information and the user-user social relationship. The model performs neighborhood feature aggregation and multi-layer information propagation on the graph structure, trying to capture social influence and collaborative signals, so as to model the expression ability of user embedding and recommendation accuracy. Unlike traditional collaborative filtering methods, GraphRec considers the bidirectional dependence of social relationships and user behavior preferences, realizes the joint learning of social information and interaction information, reduces the influence of data sparsity, and to some extent considers the interpretability of the model.

[0003] The technical measures of GraphRec are designed around three goals: "unifying multi-source graph structure", "integrating interaction and opinion information", and "characterizing the difference in social relationship strength". Specifically, it can be decomposed into 3 core operations: 1. Dual-graph modeling mechanism: In view of the limitation of traditional recommendation models that can only learn in a single interaction graph, GraphRec proposes a dual-graph modeling mechanism (Dual-Graph Modeling) that simultaneously integrates user-user social relationships and user-item interaction relationships in a unified framework. This mechanism propagates the social influence of users in the social graph and aggregates the behavior preferences between users and items in the interaction graph, thereby realizing the collaborative modeling of social and behavioral information. This method mainly fuses user features modeled in the interaction graph and the social graph. This design enables the model to jointly learn multi-source relationship information in user-user and user-item structures, thereby overcoming the shortcomings of single-graph recommendation systems that cannot fully utilize social influence.

[0004] 2. Opinion-aware aggregation mechanism: In traditional graph neural recommendation models, user-item interaction is usually simplified as a binary connection of "whether to interact", which cannot reflect the strength of user's subjective preference. GraphRec proposes an opinion-aware aggregation mechanism (Opinion-Aware Aggregation) to explicitly introduce rating semantic information in the graph convolution propagation process to realize joint learning of "interaction-opinion" two-dimensional features. The model designs an independent opinion embedding vector for each rating type r, and concatenates it with the item embedding and inputs it into the perception network: Then, the user's interaction features are obtained by a weighted aggregation function: wherein, represents the influence weight of the item on the user. This mechanism converts discrete ratings into learnable continuous representations through opinion embedding, enabling the model to distinguish between positive and negative feedback when propagating features, thus depicting more fine-grained semantic modeling in the interaction graph, to some extent improving the interpretability and stability of the recommendation results.

[0005] Attention-based Social Aggregation: To address the differences in the influence of different relationship strengths on user preferences in social networks, GraphRec designs an attention-based social aggregation mechanism. In the user-user social graph, an attention allocation mechanism is introduced to realize dynamic modeling of relationship weights. Specifically, the model calculates the social attention weight between the target user and its neighbor users through a double-layer neural network and uses Softmax normalization to ensure that the sum of all weights is 1. The aggregation process is represented as: wherein, represents the interaction features of the neighbor users, and the social edge influence strength.

[0006] This mechanism enhances the recommendation system's ability to perceive complex social signals to some extent. However, the above method usually models based on the original social structure, does not remove noise connections at the structural level in the social graph, and lacks a joint optimization mechanism for popularity bias and long-tail preference in the recommendation results.

[0007] Literature: Denoised Self-Augmented Learning for Social Recommendation DSL (Denoised Self-Augmented Learning) is a self-supervised learning framework for social recommendation tasks. Its core idea is to introduce denoising and cross-view enhancement mechanisms in user social relationships. By filtering low-confidence social connections and aligning the feature representations of the interaction and social views, the model's robustness to social noise and consistency in user representation are improved. This method to some extent considers the noise interference and feature deviation problems in social data, and is suitable for recommendation scenarios with complex social relationships and sparse interaction data.

[0008] The technical design of DSL revolves around three core objectives: "social relationship denoising", "cross-view feature alignment", and "self-supervised enhancement". It models and optimizes the problems of large social relationship noise and sparse interaction data in social recommendation tasks. Specifically, it can be decomposed into two key operations: Denoising self-enhancement learning framework: DSL proposes a denoising self-enhancement learning framework to alleviate the representation bias problem caused by noisy social relationships in social recommendation. The framework introduces a learnable edge confidence weight in the social view to suppress unreliable or interest-independent social connections, thereby obtaining purer social signals. Specifically, the model assigns weights to social edges based on the similarity information of users in the interaction view and performs denoising aggregation during graph convolution propagation: where denotes the social neighbor set of user , and denotes the confidence of edge (i,j). This mechanism attempts to achieve adaptive denoising of social graph structure, allowing the model to preferentially aggregate high-confidence neighbor information during feature propagation, thereby mitigating the interference of social noise and providing a stable representation basis for subsequent cross-view feature alignment.

[0009] Cross-view representation alignment mechanism: To unify the embedding distribution of users in the social view and the interaction view, DSL designs a cross-view representation alignment mechanism. The model learns two sets of embeddings and on the user-item interaction graph and the user-social relationship graph respectively, and introduces a self-supervised consistency constraint to ensure that the representation of the same user in both views remains semantically consistent. This mechanism unifies the social and interaction feature spaces under the constraint of self-supervised signals, promoting semantic alignment and information complementarity between the two views, and attempting to enable the model to more fully utilize multi-source relationship structures for recommendation modeling.

[0010] In existing social recommendation techniques, the modeling process of social graphs and interaction graphs is often handled separately, making it difficult to achieve collaborative optimization between the two. On the one hand, traditional social recommendation models generally assume that "social relationships are similar in interest", but this homogeneity assumption does not always hold. When there are interest conflicts or noise relationships in social connections, the information propagated along non-homogeneous edges will introduce noise, causing user embedding deviation and a decrease in the reliability of social signals. On the other hand, existing denoising strategies (such as DSL) are mostly based on heuristic rules, lack consistency with personalized recommendation goals, and are difficult to adaptively distinguish between "valuable connections" and "redundant relationships" in complex social structures. In addition, mainstream attention or pruning methods only focus on local similarity weights, ignoring the joint influence of user behavior similarity and graph structure centrality, making it difficult to accurately identify trusted social connections. At the same time, long-tail users and niche items are still underestimated, and the embedding space dominated by popular users and high-frequency interactions is obscured, resulting in insufficient diversity and fairness in the recommendation results.

[0011] To solve the above problems, the present application proposes a social recommendation method based on denoising and preference balancing, which realizes the adaptive removal of noise social relationships and the balanced modeling of user preferences through a structure-aware social graph pruning mechanism and a long-tail-popular preference objective function. Among them, the structure-aware social graph pruning module evaluates the reliability of social connections based on behavior similarity and node centrality, adaptively removes noise edges, and improves the reliability of social signal propagation; the long-tail-popular preference objective function balances the exposure probability of long-tail items and popular items through dynamic weights, alleviates recommendation bias, and enhances the diversity and fairness of the results. The synergistic effect of the two makes the model more robust, balanced and trustworthy in personalized recommendation in a noisy, sparse and structurally heterogeneous social environment.

[0012] In the field of social recommendation modeling, GraphRec was the first to realize the joint modeling of social relationships and user-item interaction graphs, but its core framework based on static social connections and linear feature fusion has multiple limitations in dealing with complex features such as noise propagation, hierarchical structure differences and long-tail bias in social graphs, making it difficult to fully exploit high-confidence information in social relationships and the diversity of user preferences. This is manifested in the following three aspects: I. Lack of adaptability to social relationship noise and strong and weak heterogeneity Real-world social networks simultaneously exhibit strong relationships (close friends, core colleagues) and weak relationships (general followers, low-interaction fans), and even pseudo-relationships (zombie followers, mutual following among marketing accounts). These edges do not necessarily reflect preference similarity. GraphRec learns latent factors on the user side through "opinion-aware interaction representation + adjacency aggregation / attention aggregation," but it assumes neighbor reliability and propagates layer by layer along the edges: mean aggregation treats weak edges and noisy edges with equal weight, and although attention aggregation can learn weights, it still scores based on the target user and interaction vector after encoding, lacking pre-encoding structural denoising and hop credibility constraints. Noise signals are amplified during multi-hop propagation, ultimately causing user representations to be influenced by non-like-minded neighbors, resulting in a mixed embedding distribution, and making it more prone to over-smoothing in sparse scenarios. Correspondingly, this invention introduces structure-aware social graph pruning before encoding: on the one hand, it uses behavioral similarity (common interaction, similarity measurement) to characterize "whether they share the same interests," and on the other hand, it uses structural centrality (degree) to characterize "connection credibility / propagation value." These two are combined into a learnable edge reliability score, and low-confidence edges are either softly weakened or hard-removed, retaining only high-quality social signals for message passing. Simultaneously, consistent penalties / regularizations are maintained across layers in subsequent propagation to suppress noise accumulation. This mechanism significantly reduces interference from weak / pseudo-relationships without disrupting strong relationship representation, allowing dual-graph fusion to focus more on genuine social dependencies. This results in more robust user representations and improved downstream recommendation performance in complex, noisy, and sparse social environments.

[0013] II. Insufficient adaptability to the integration layer of social and interactive views. In social recommendation tasks, user behavior is typically driven by both social relationships and individual interests: the social view reflects long-term group attachment and relationship influence, while the interaction view reflects short-term individual preferences and behavioral dynamics. An ideal social recommendation model should achieve hierarchical and dynamic fusion between these two types of views, ensuring that user embeddings maintain both social consistency and individual differences. However, GraphRec employs a parallel strategy of independent encoding and linear concatenation in its modeling process, learning social and interaction embeddings separately before weighted summation, lacking cross-view structural alignment and dynamic weight adjustment mechanisms. This static fusion approach exhibits significant limitations in capturing multi-layered semantic information: when social influence is strong, interaction features are easily over-smoothed, leading to convergent user preference expressions; while when individual behavior is sparse, social embeddings struggle to provide sufficient distinctiveness, resulting in ambiguous user representations after fusion. Furthermore, GraphRec's fixed weighting method cannot adapt to the non-uniformity of social relationships across different circles, making it difficult for the model to distinguish between strong relationship constraints and weak relationship noise, thus weakening the effective fusion of cross-view features. To address the aforementioned issues, this invention introduces a cross-view adaptive alignment mechanism after structure-aware pruning in dual-graph joint modeling, and designs a gated-attention joint fusion module: During the feature propagation stage, the gated unit adaptively adjusts the information flow ratio based on the confidence level of social and interaction signals, effectively suppressing interference from noisy social edges; during the fusion stage, the global attention module dynamically measures the importance of social and interaction features using learnable weights, achieving adaptive alignment and weighted integration at the semantic layer. Through this mechanism, DBA-SR can achieve hierarchical information interaction and dynamic feature fusion on high-confidence social structures, enabling user embeddings to possess both social stability and individual differences, thereby generating more hierarchical and interpretable preference representations in complex and heterogeneous social environments.

[0014] III. Insufficient Modeling Capabilities for Long-Tail Users and Niche Items Real-world social recommendation systems commonly suffer from popularity bias and long-tail imbalance: a few high-frequency users and popular items are over-emphasized due to frequent interactions and dense social connections, while most low-activity users and niche items are ignored due to sparse signals, resulting in concentrated and insufficiently personalized recommendation results. GraphRec, in its unified embedding update, does not hierarchically model the differences between activity and popularity, allowing popular nodes to dominate aggregation and making it difficult for long-tail features to learn. Furthermore, it lacks joint constraints on social propagation and behavior frequency, failing to suppress the popularity amplification effect and reducing diversity and fairness. To address this, this invention introduces a long-tail-popularity preference balancing objective function to model the popularity distribution of users and items, dynamically increasing the weight of long-tail samples. Combined with a cross-view attention fusion mechanism, it adaptively controls the exposure ratio of social and interaction features, ensuring that low-frequency nodes receive sufficient updates. This mechanism alleviates the excessive propagation of popular information from both structural and optimization perspectives, achieving an overall balance between accuracy, diversity, and fairness.

[0015] The optimization goal of DSL revolves around "denoising alignment and self-supervised enhancement between social and interactive views." The model performance is highly dependent on the quality of noise modeling and view alignment, and there are two main problems: I. The Staticity of Noise Modeling and Over-reliance on Confidence Weights While DSL introduces a social denoising module to mitigate the interference of low-confidence social relationships in social recommendation, its denoising mechanism relies on static confidence weights and lacks the ability to dynamically adjust them during training. When noise in social relationships changes over time or behavioral patterns, fixed weights may cause the model to prematurely abandon potentially effective edges, thereby weakening the coverage of social propagation. For example, for user pairs that "interact occasionally but have similar interests," DSL's confidence estimation tends to underestimate their relationship value, causing useful social signals to be incorrectly truncated, ultimately limiting the model's accuracy in characterizing users' true preferences.

[0016] II. The unidirectional nature and insufficient information coupling of cross-view alignment DSL achieves consistency in social and interactive embeddings through cross-view alignment, but its alignment process is mainly based on unidirectional semantic similarity constraints, failing to explicitly model the bidirectional interaction between social and behavioral features. This unidirectional alignment easily leads to embedding space collapse, with social information excessively dominating user features, while individual behavioral signals struggle to feed back into social view updates. For example, in multi-interest user scenarios, different interest dimensions are uniformly mapped to a single social feature center, causing the disappearance of individual differences and the ambiguity of preference expression, thus reducing adaptability in diverse recommendation scenarios.

[0017] Deficiencies to be overcome / problems to be solved: Social relationship noise is not effectively identified and suppressed: Existing methods generally rely on static graph structures when modeling social relationships, which cannot distinguish between reliable and redundant edges, resulting in social propagation being interfered with by false relationships and reducing the credibility of feature representation.

[0018] Uneven preference distribution leads to unfair recommendations: In social recommendation scenarios, popular users and items have too much weight, while long-tail users and niche interests are ignored, making it difficult to balance performance and diversity.

[0019] Insufficient integration of social and interaction graphs: Existing models mostly process social and interaction graphs independently, without establishing an adaptive fusion mechanism based on behavioral similarity and structural centrality, making it difficult to capture multi-level association patterns of user preferences. Summary of the Invention

[0020] The purpose of this invention is to provide a social recommendation method and system based on denoising and preference balancing, aiming to solve problems such as insufficient filtering of noisy social relationships, imbalance between individual preferences and group popularity, and insufficient integration of social and interactive information in existing social recommendation methods. This invention constructs a social recommendation method based on denoising and preference balancing (DBA-SR), which uses adaptive graph structure optimization and long-tail preference constraint mechanisms to achieve accurate modeling of real social relationships and balanced expression of diverse preferences, thereby improving the reliability, fairness, and recommendation accuracy of the model in noisy and sparse social environments.

[0021] To address the aforementioned technical problems, this invention provides a social recommendation method based on denoising and preference balancing, comprising the following steps: Building a social graph; Denoise the social graph to obtain the denoised social graph; Using the normalized user item interaction matrix and the denoised social graph as input, probabilistic user embeddings are learned through two-branch variational encoding, and then user fusion representations are obtained through variational regularization, gating denoising and attention fusion. Based on user fusion representation, the interaction graph and social graph are jointly reconstructed, and the training target is obtained by optimizing the objective through preference balancing.

[0022] Preferably, the social graph is denoised to obtain a denoised social graph, specifically including the following steps: Calculate the edge similarity of the social graph to obtain a similarity score; The social graph is denoised based on the similarity score to obtain the denoised social graph.

[0023] Preferably, the formula for calculating the similarity score is: in: Represents the similarity score. and Representing users respectively and A collection of interactive items, where deg(v) represents a node. The degree, It is a constant; Based on the similarity score, the social graph is denoised to obtain the denoised social graph. The specific steps include: Based on the set of similarity scores for all edges Calculate the distribution sensitivity threshold based on the mean and standard deviation: in: Indicates the distribution-sensitive threshold. and Let represent the mean and standard deviation of the similarity scores, respectively. Indicates the intensity of pruning; Introduce a quantile threshold to measure the sparsity of the overall edge distribution: in: Represents the sparsity threshold. Indicates the quantile threshold; The maximum value of the distribution sensitivity threshold and the sparsity threshold is used as the unified pruning threshold: When the edge weight score If the value is greater than θ, the edge is retained; otherwise, it is removed, thereby obtaining the denoised adjacency matrix and the denoised social graph. Symmetrical normalization is performed on the denoised social graph: in: This represents a social graph after noise reduction.

[0024] Preferably, the normalized user item interaction matrix and the denoised social graph are used as inputs. Probabilistic user embeddings are learned through two-branch variational encoding, and then user fusion representations are obtained through variational regularization, gating denoising, and attention fusion. Specifically, the following steps are included: Normalized user project interaction matrix With denoised social graph As input, probabilistic user embedding representations are extracted using two parallel variational encoders; these user embedding representations include social embeddings. and interaction representation ; Kullback–Leibler divergence is introduced as a variational regularization term to calculate the distance between the posterior distributions of behavioral and social branches and the standard Gaussian prior. Introducing content-based gating mechanisms for social embedding Perform filtering, combined with interactive representation Generate a gating vector, and obtain the gating social representation based on the gating vector; An attention fusion module is introduced to adaptively balance the importance of behavioral embedding and social embedding, and user fusion representation is obtained based on the gated social representation.

[0025] Preferably, a normalized user item interaction matrix is ​​used. With denoised social graph As input, probabilistic user embedding representations are extracted using two parallel variational encoders, specifically including the following steps: In each branch, the initial adjacency matrix of the interactive branches is set to Social branches are set as follows And the node representation is updated layer by layer through L-layer nonlinear propagation: The final output high-level features and The parameters are used to generate the Gaussian distribution, thus obtaining a probabilistic user representation: in, and These represent the mean vector and standard deviation vector of the latent representation of users from the perspective of interaction behavior, respectively. and Let these be the mean vector and standard deviation vector of the latent representation of a user from the perspective of social relationships, respectively. The distribution is then sampled using a reparameterization technique: in, and A random vector that follows a standard normal distribution is used to introduce randomness to characterize the uncertainty of the user's potential representation; the symbol ⊙ represents element-wise multiplication.

[0026] Preferably, the formula for calculating the distance between the posterior distributions of the behavioral branch and the social branch and the standard Gaussian prior is as follows: In the formula: This represents the distance between the posterior distributions of the behavioral and social branches and the standard Gaussian prior. The formula for calculating the gate vector is: in, and Let represent the trainable weight matrices used for linear mapping of user interaction embeddings and social embeddings, respectively. This represents the corresponding bias vector; This represents a non-linear activation function used to restrict the values ​​of the gating vector within a preset range, thereby achieving soft selection of social information; The formula for calculating the gated social representation is as follows: Among them, symbols This indicates element-wise multiplication. This represents the trainable weight matrix used for feature enhancement of social embeddings. This represents the corresponding bias vector; the hyperbolic tangent function is used to introduce nonlinear mapping capabilities to enhance the expressive power of social features. Attention weights are concatenated from behavioral and social features: in, This represents a trainable weight matrix used to perform feature transformation on the concatenated joint features, which is used to capture the correlation between user behavior features and social features; This represents a trainable weight vector used to map the transformed features to attention scores; This represents a non-linear activation function used to normalize or range-constrain attention scores; (symbol) This represents a vector concatenation operation; The user fusion representation is as follows: in, This represents the user fusion representation.

[0027] Preferably, based on user fusion representation, joint reconstruction of interaction graph and social graph is performed, and the training objective is obtained through preference balancing optimization. Specifically, this includes the following steps: Reconstruct the interaction and social relationships to obtain social intimacy; Based on a dual-objective optimization mechanism, we obtain long-tail preference enhancement and popularity suppression; By constraining the consistency between the predicted social intimacy and the denoised social relationship through cross-entropy loss, a consistency constraint term is obtained; The training objective is obtained based on the long-tail preference enhancement, popularity suppression, and consistency constraint terms.

[0028] Preferably, the interaction and social relationships are reconstructed to obtain social intimacy, specifically including the following steps: Leveraging user embedding Embedded with items Calculate interaction preference scores to predict the strength of the association between users and items: in, This represents the item embedding representation used for reconstructing interaction relationships, which characterizes the latent semantic features of items in the prediction space; Using the same user embedding, social intimacy is calculated via a multilayer perceptron to recover the latent structural relationships in the denoised social graph: in, This represents the predicted social intimacy between user u and user v, used to characterize the strength of their social association in the latent feature space; and These represent the user embedding representations corresponding to user u and user v, respectively, and the user embedding representations serve as the input to the social decoding module; The predicted social relationship matrix is ​​obtained by summing up the predicted social intimacy values ​​of all users for (u,v).

[0029] Preferably, based on a dual-objective optimization mechanism, long-tail preference enhancement and popularity suppression are obtained, specifically including the following steps: Increase the learning signal for rare items through a reverse popularity weighting mechanism: in, This represents the popularity of item i, used to characterize the frequency of the item's appearance in historical interaction data; This represents the long-tail enhancement coefficient, used to adjust the enhancement strength of the anti-prevalence weighting mechanism. Its value controls the extent to which the weight of low-prevalence items is increased. Subsequently, the reweighted binary cross-entropy loss function is defined as follows: Introducing a popularity regularization term suppresses overfitting of high-popularity samples and promotes a more balanced recommendation distribution: Where λ>1 is used to control the steepness of the penalty term; Interactive reconstruction targets integrate two types of mechanisms: Among them, coefficient The balance between enhancement and inhibition is controlled; The formula for calculating the consistency constraint is as follows: The formula for calculating the training objective is: Where β represents the weight coefficient of the KL divergence regularization term, which is used to adjust the influence of probability distribution constraints on the overall training objective.

[0030] This invention also provides a social recommendation system based on denoising and preference balancing, comprising: The social graph building module is used to build social graphs; The structure-aware social graph pruning module is used to denoise the social graph and obtain a denoised social graph. The dual variational coding and social representation fusion module is used to learn probabilistic user embeddings through dual-branch variational coding, and then obtain the user fusion representation through variational regularization, gating denoising and attention fusion. The dual-graph decoding and preference balance optimization module is used to jointly reconstruct the interaction graph and social graph based on the user fusion representation, and obtain the training target through the preference balance optimization objective.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: I. Advantages of Structure-Aware Social Graph Denoising: Enhanced Relationship Credibility and Embedding Robustness From a theoretical perspective, real social relationships in social recommendations are often mixed with noisy edges and weakly related edges. Direct propagation of these edges can amplify misleading information and cause embedding distortion. The structure-aware social graph pruning mechanism introduced by DBA-SR adaptively evaluates edge quality using both behavioral similarity and structural centrality metrics, achieving dynamic filtering of noisy edges and enhanced propagation of social signals. This method avoids the coarse operations of traditional static threshold-based or heuristic pruning, ensuring that the retained social connections better reflect real behavioral relevance and contributing to improved stability of the denoised social graph in terms of user similarity distribution.

[0032] II. Advantages of Preference Balance Optimization Objectives: It balances performance, diversity, and fairness. Existing recommendation models are generally affected by "popularity bias," with popular items being overexposed and long-tail items struggling to be recommended. DBA-SR introduces a long-tail-popularity-aware preference objective function in the decoding stage. It uses inverse popularity weighting and exposure suppression regularization for joint optimization to adjust the exposure distribution of popular and long-tail items, achieving a balanced preference model.

[0033] III. Advantages of Dual Variational Coding Fusion: Unified Modeling of Uncertainty Awareness and Social Feature Fusion Traditional social recommendation models often rely solely on fixed embeddings, making it difficult to represent the potential uncertainties of behavioral and social features, thus weakening individual differences. DBA-SR, through a dual variational coding structure, learns latent distribution parameters in parallel on the interaction graph and the denoised social graph to obtain probabilistic user embedding representations. This mechanism can explicitly model the randomness of user behavior differences and changes in the social environment, which is beneficial for enhancing the generalization ability and stability of the embedding representation under different user behavior scenarios. Simultaneously, an attention-gated fusion mechanism is introduced to dynamically allocate weights between behavioral and social embeddings, achieving adaptive fusion of the two types of features, ensuring that the user representation retains both individual behavioral characteristics and social consistency. Attached Figure Description

[0034] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] Figure 1 This is a schematic diagram of the overall structure of the DBA-SR framework; Left: A structure-aware social graph denoising module that adaptively removes noise or weak social connections; Middle: A dual variational coding module for learning user representations with uncertainty awareness; Right: An attention-gated fusion mechanism and a dual-graph decoding module responsible for relationship reconstruction and mitigating exposure bias. Red edges represent potential connections newly inferred by the model during training. The entire framework achieves collaborative learning of social graph structure reconstruction, variational regularization, and interaction reconstruction objectives with bias-aware constraints through joint optimization.

[0036] Figure 2 This is a diagram of a gating mechanism; Figure 3 This is a schematic diagram of the attention fusion mechanism. Detailed Implementation

[0037] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0038] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0039] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0040] The present invention will now be described in further detail with reference to the accompanying drawings: To better illustrate the technical effects of the present invention, the present invention provides the following specific embodiments to illustrate the above technical process: Example 1: A social recommendation method based on denoising and preference balancing is a social recommendation method based on structure-aware denoising, variational coding fusion and preference balancing optimization. It is applicable to personalized recommendation tasks in various scenarios such as e-commerce recommendation, social media content recommendation and interest community recommendation.

[0041] The overall technical solution of this invention includes three core stages: social graph structure denoising, variational coding and social representation fusion, and dual-graph reconstruction optimization that integrates long-tail and popularity constraints. Each stage is sequentially connected and works collaboratively. The specific implementation process is as follows: 1. Structure-aware social graph pruning This step aims to utilize the behavioral similarity and topological centrality between nodes to calculate the quality of social edges, adaptively remove noisy connections while preserving high-confidence relationships, thereby generating a sparser and semantically reliable social graph structure. Specifically, it includes: 1.1 Social Graph Modeling First, construct the social graph. ,in Represents the set of all user nodes. This represents the set of social edges between users (such as friend relationships and follow relationships).

[0042] In the formula: Indicates user node 1~ N ; , Indicates the first One user node.

[0043] 1.2 Social Edge Similarity Calculation After constructing the social graph, a similarity score is calculated for each social edge (u,v) to measure the strength of relationships between users. This score considers both user behavior similarity and structural centrality: the former measures the similarity of users across sets of interacting items using the Jaccard index, while the latter reflects a user's importance within the social network using a normalized value of node degree. Its definition is as follows: in: and Representing users respectively and A collection of interactive items, where deg(v) represents a node. The degree, To prevent small constants with a denominator of zero, this indicator comprehensively captures the behavioral similarity and structural weights of social relationships, providing a quantitative basis for subsequent noise edge pruning.

[0044] 1.3 Hybrid Criterion Pruning and Social Graph Denoising To eliminate noisy edges and preserve high-quality connections in social graphs, a hybrid pruning mechanism integrating statistical and quantile features is constructed. First, it is based on the similarity score set of all edges. Calculate the distribution sensitivity threshold based on the mean and standard deviation. in, Indicates the distribution-sensitive threshold. and Let represent the mean and standard deviation of the similarity scores, respectively. Control the intensity of pruning.

[0045] At the same time, a quantile threshold is introduced to measure the sparsity threshold of the overall edge distribution: in: Indicates sparsity. Indicates the quantile threshold; The maximum value of the two values ​​is ultimately used as the unified pruning threshold. When the edge weight score If the value is greater than θ, the edge is retained; otherwise, it is pruned, thus obtaining the denoised adjacency matrix. Subsequently, to adapt to subsequent graph convolution calculations, the following steps are performed: After denoising, the social graph is symmetrically normalized. in: This represents the adjacency matrix of the social graph after denoising and symmetric normalization.

[0046] This process combines global variance and distribution sparsity to effectively filter weak correlations and redundant edge connections, improving the reliability of social signal propagation and providing a clean and stable structural foundation for subsequent user representation learning.

[0047] 2. Fusion of Dual Variational Coding and Social Representation This step aims to leverage a dual variational coding framework and a gated attention fusion mechanism to jointly model the uncertainty perception embedding learned by users in the social graph with behavioral features, thereby generating a unified user representation with social context semantics. Specific implementation: 2.1 Dual Variational User Representation Coding To fully utilize the denoised social graph structure optimized by structure-aware pruning, this module designs a two-branch variational coding framework to jointly model the uncertainty of user representation from both user interaction behavior and social relationship structure perspectives. The model uses a normalized user-item interaction matrix. With denoised social graph As input, probabilistic user embedding representations are extracted through two parallel variational encoders.

[0048] Among them, the normalized user-project interaction matrix The acquisition process includes: (1) Construct a user-project interaction graph based on the historical interaction records between users and projects, where the interaction value is set to 1 when there is interaction between the user and the project, and 0 otherwise; (2) Construct the user-item interaction matrix R based on the user-item interaction graph; (3) Normalize the user-item interaction matrix R according to the user node degree and the item node degree to obtain the normalized user-item interaction matrix. .

[0049] The denoised and normalized social relationship matrix The acquisition process includes: (1) Construct a user-user social graph based on the original social relationships between users; (2) Perform structure-aware pruning on the social edges in the user-user social graph to remove weak or noisy social connections and obtain a denoised social relationship graph. (3) Construct a social relationship matrix based on the denoised social relationship graph and normalize it to obtain a normalized social relationship matrix. .

[0050] Obtained through the above methods and These are used as inputs for user interaction behavior information and social relationship structure information, respectively, in the subsequent two-branch variational coding process, thereby achieving joint modeling of the uncertainty of user representation.

[0051] In each branch, the initial adjacency matrix of the interactive branches is set to Social branches are set as follows And the node representation is updated layer by layer through L-layer nonlinear propagation: The final output high-level features This represents the user node feature matrix obtained after the interaction branch propagates at layer l. This represents the user node feature matrix obtained after the social branch propagates at layer l. Both are used to generate the parameters of the Gaussian distribution, thus obtaining the probabilistic user representation: in, and These represent the mean vector and standard deviation vector of the latent representation of users from the perspective of interaction behavior, respectively. and Let represent the mean vector and standard deviation vector of the latent representation of a user from the perspective of social relationships, respectively.

[0052] Subsequently, the distribution was sampled using reparameterization techniques to obtain the social embeddings. and interaction representation : in, and A random vector that follows a standard normal distribution is used to introduce randomness to characterize the uncertainty of the user's potential representation; the symbol ⊙ represents element-wise multiplication.

[0053] This dual variational mechanism enables the model to simultaneously learn user interaction preferences and social feature distributions within an uncertainty framework, laying the foundation for subsequent multi-source feature fusion and robust recommendation modeling.

[0054] 2.2 Variational Regularization To constrain the distribution characteristics of the latent space and prevent overfitting, Kullback–Leibler (KL) divergence is introduced as a variational regularization term to calculate the distance between the posterior distributions of the behavioral and social branches and the standard Gaussian priors: Through this regularization process, the model can learn decoupled and uncertainty-aware user representations based on the denoised graph structure, providing a more stable latent feature space for subsequent recommendation tasks.

[0055] 2.3 Gated Social Network Noise Reduction To mitigate noise interference in social connections, a content-based gating mechanism is introduced to control social embeddings. Perform filtering, combined with interactive representation Generate gated vectors: in, and Let represent the trainable weight matrices used for linear mapping of user interaction embeddings and social embeddings, respectively. This represents the corresponding bias vector; This represents a non-linear activation function used to restrict the values ​​of the gating vector within a preset range, thereby achieving soft selection of social information.

[0056] Then, gating filtering is applied to the social embedding representation to obtain the denoised social feature representation. The calculation method is as follows: Among them, symbols This indicates element-wise multiplication. This represents the trainable weight matrix used for feature enhancement of social embeddings. This represents the corresponding bias vector; the hyperbolic tangent function is used to introduce nonlinear mapping capabilities to enhance the expressive power of social features.

[0057] This mechanism can adaptively retain relevant social information and suppress invalid connections, thereby generating more robust social embeddings.

[0058] 2.4 Attention-based behavior – social integration In obtaining denoised social representations Subsequently, the model introduces an attention fusion module to adaptively balance the importance of behavioral and social embeddings. The attention weights are calculated by concatenating behavioral and social features: in, This represents a trainable weight matrix used to perform feature transformation on the concatenated joint features, which is used to capture the correlation between user behavior features and social features; This represents a trainable weight vector used to map the transformed features to attention scores; This represents a non-linear activation function used to normalize or range-constrain attention scores; (symbol) This indicates a vector concatenation operation.

[0059] Through the above attention weights It can adaptively measure the relative importance of user behavior information and social information in the current user representation, thereby providing a basis for subsequent feature fusion.

[0060] The final user integration characteristic is: This dynamic fusion strategy enables the model to dynamically adjust the weights of behavior and social information based on the context, thereby generating a unified user representation with social awareness and personalized features, providing robust input for subsequent preference prediction.

[0061] 3. Dual-graph decoding and preference balancing optimization This phase, based on the fusion-based uncertain user representation obtained in the previous phase, completes the joint reconstruction of the user-item interaction graph and the social graph. It also mitigates popularity bias and uneven long-tail exposure issues through preference balancing optimization. Specifically, the following is implemented: 3.1 Reconstruction of the Dual-Graph Structure To maintain semantic consistency and structural coherence, the model constructs a dual-decoder framework to reconstruct interaction relationships and social relationships separately. This is achieved by utilizing fused user embeddings. Embedded with items The interaction preference score is used to predict the strength of the association between users and items, as calculated below: in, This represents the item embedding representation used for reconstructing interaction relationships, which characterizes the latent semantic features of items in the prediction space.

[0062] Social graph decoding: Utilizing the same user embedding to calculate social intimacy through a multilayer perceptron (MLP) to recover the latent structural relationships in the denoised social graph.

[0063] in, This represents the predicted social intimacy between user u and user v, used to characterize the strength of their social association in the latent feature space; and These represent the user embedding representations corresponding to user u and user v, respectively, and the user embedding representations serve as the input to the social decoding module; The predicted social relationship matrix is ​​obtained by summing up the predicted social intimacy values ​​of all users for (u,v).

[0064] Through a dual-branch reconstruction mechanism, the model maintains semantic alignment and structural consistency in both the interaction and social domains, ensuring the transferability of user embeddings across different graph structures.

[0065] 3.2 Long Tail and Popularity Perception Preference Optimization Objectives To address the issues of uneven exposure and popularity bias in the recommendation process, a dual-objective optimization mechanism was designed: 3.2.1 Enhanced Long-Tail Preference Increase the learning signal for rare items through a reverse popularity weighting mechanism: in, This represents the popularity of item i, used to characterize the frequency of the item's appearance in historical interaction data; This represents the long-tail enhancement coefficient, used to adjust the enhancement strength of the anti-popularity weighting mechanism. Its value controls the extent to which the weight of low-popularity items is increased.

[0066] Subsequently, the reweighted binary cross-entropy loss function is defined as follows: By adaptively amplifying the learned signals of low-frequency interactive items, this mechanism can improve the recall rate of long-tail items and enhance the diversity of recommendation results.

[0067] 3.2.2 Prevalence Suppression Regularization To prevent highly popular items from dominating the prediction results, a popularity regularization term is introduced to suppress overfitting of highly popular samples and promote the balance of the recommendation distribution. The calculation is as follows: Here, λ>1 is used to control the steepness of the penalty term. A larger pi value will bring a stronger regularization effect, thereby suppressing excessively high relevance scores for popular items and promoting fair exposure of recommendation results in the overall item space.

[0068] The ultimate goal of interaction reconstruction is to integrate two types of mechanisms: Among them, coefficient By controlling the balance between enhancement and suppression, optimization can be achieved that balances performance and fairness.

[0069] 3.3 Social Relationship Consistency Constraints In the social relationship reconstruction phase, the model uses cross-entropy loss to constrain the predicted social intimacy. Social relationships after noise reduction Consistency: in, Social connections derived from structure-aware denoising and meeting a high confidence threshold are used as weak supervision signals. This constraint ensures semantic consistency in the social domain, prevents embedding drift caused by variational sampling, and ensures that user representations possess coherent relational information in both the behavioral and social domains.

[0070] 3.4 Joint Optimization Objective The final training objective incorporates three types of constraints: Where β represents the weight coefficient of the KL divergence regularization term, which is used to adjust the influence of probability distribution constraints on the overall training objective.

[0071] Advantages of this invention: 1. Denoising of structure-aware social graphs, enhancement of social signals, and improvement of representation robustness. This invention calculates behavioral similarity based on the overlap of user interactions (the set of purchased items) and combines node degree centrality to reflect social influence, thus obtaining edge weights. Subsequently, a combined standard deviation threshold and quantile threshold screening strategy was used to adaptively remove low-quality or noisy edges, resulting in a denoised social graph with optimized structure.

[0072] Traditional social recommendation models typically rely directly on existing social relationships for propagation, making them susceptible to interference from "non-homogeneous relationships" and "weak connections." The structure-aware denoising strategy of this invention effectively filters out invalid or spurious social edges, strengthens the expression of social signals, and significantly improves the robustness and generalization of user representations.

[0073] 2. Preference balance optimization objective: joint equilibrium of long-tail compensation and popularity suppression This invention designs a preference balance optimization objective, employing a dual mechanism of long-tail reinforcement and popularity constraint to jointly optimize user-item preference prediction. Specifically, it calculates normalized popularity based on item interaction frequency. Assigning countercurrent weights to sparse long-tail samples Adaptively enhance the learning signal of low-exposure items; simultaneously, introduce a popularity penalty term. Regularization constraints are applied to overconfidence predictions of popular items, thereby achieving bidirectional balance optimization in the overall loss.

[0074] Traditional recommendation models tend to overemphasize high-frequency, popular items while neglecting long-tail items, leading to a decline in recommendation diversity and fairness. This method employs adaptive reweighting and regularized balancing to compensate for the exposure of long-tail items and suppress the confidence of popular items, thereby improving the coverage and personalized fairness of recommendation results while maintaining performance.

[0075] 3. Dual variational coding fusion, synergistic enhancement of uncertainty modeling and social behavior. This invention proposes a dual variational encoding fusion mechanism, which uses the interaction graph and the denoised social graph as parallel inputs to a dual-branch variational encoder, and models the uncertainty of user behavior and social features through a Gaussian distribution. After generating probabilistic embeddings, a gating mechanism is introduced to filter noisy social information, and attention fusion is combined to adaptively balance the weights of social and behavioral features, achieving dynamic coordination between social semantics and behavioral preferences.

[0076] Traditional models struggle to handle noisy social connections and sparse interactions, often leading to feature overfitting or weight imbalance. This method captures uncertain feature distributions through dual variational coding, and combines gating and attention fusion to achieve signal enhancement and noise suppression, ensuring the model remains robust and stable in complex social environments.

[0077] Example 1: Experimental dataset: Douban-book: This dataset contains data on relationships between users on the Douban Books platform, as well as data on their book preferences.

[0078] Douban-movie: This dataset contains social network data and movie rating records collected from a group of users on the Douban Movie platform.

[0079] Yelp: This dataset contains social connections established by users on the Yelp (a well-known US business review platform) review platform, along with their rating data. It provides a valuable foundation for analyzing user preferences and behaviors in personalized recommendation scenarios.

[0080] Evaluation indicators: Following common practice, this invention uniformly transforms user-item interactions in each dataset into an implicit feedback form: if user u interacts with item i, it is recorded as rui=1; otherwise, rui=0. Each dataset is then randomly partitioned, with 80% of the observed interactions used for training and 20% for testing; this process remains consistent across all datasets. To measure recommendation performance, this invention employs two commonly used industry metrics: Recall@N and NDCG@N. Recall@N evaluates the proportion of relevant items successfully recalled by the model in the first N recommendations, reflecting recall capability; NDCG@N (Normalized Decay Cumulative Gain) considers both relevance and ranking position, giving higher weight to correct recommendations with higher rankings, and is used to evaluate ranking quality and user experience.

[0081] To verify the effectiveness and generalization ability of the proposed method, extensive experiments were conducted on three benchmark datasets: Douban-Book, Douban-Movie, and Yelp. DBA-SR was also compared with several state-of-the-art baseline methods. Experimental results are summarized in Table I. The baseline models involved are described below: MF: A matrix factorization method that decomposes the user-item interaction matrix into low-rank latent representations of users and items, and introduces bias terms and regularization for preference prediction.

[0082] SocialMF: A method based on social networking that incorporates social network structures into the model by introducing regularization terms, encouraging users with connected social relationships to learn similar latent representations.

[0083] DiffNet: A social recommendation model that spreads user preferences across a social graph through an influence diffusion mechanism, while capturing both direct and higher-order social influence effects.

[0084] NGCF: A collaborative filtering model based on graph neural networks that models higher-order connectivity and collaborative signals by propagating embedded representations on the user-item interaction graph.

[0085] LightGCN: A simplified GCN model for collaborative filtering that learns cleaner user-item embeddings using only linear neighborhood aggregation, removing feature transformations and non-linear activation functions.

[0086] MHCN: A social recommendation model that uses multi-channel hypergraph convolution to characterize diverse high-order social relationships and user preferences in different social contexts.

[0087] SEPT: A self-supervised socially-aware recommendation model based on a tri-training framework of multi-graph encoders, which fully utilizes social information to enhance representation learning capabilities.

[0088] DcRec: A recommendation framework that decouples user representations from the item and social domains through contrastive learning, and improves recommendation performance through cross-domain knowledge transfer.

[0089] SimGCL: A recommendation model based on contrastive learning that reduces popularity bias and improves training efficiency by removing complex graph augmentation operations and introducing uniform noise to generate contrastive views.

[0090] XSimGCL: A contrastive learning recommendation model that also constructs a contrastive view by removing graph augmentation and introducing uniform noise, thereby improving model efficiency while reducing popularity bias.

[0091] DVGRL: A dual variational graph reconstruction framework for social recommendation, which jointly models user interaction graphs and social graphs, reconstructs them through variational distributions, and utilizes cross-domain fusion to achieve knowledge transfer.

[0092] Table I compares recommendation performance on the three datasets. Overall, the method of this invention consistently outperforms all comparable methods on all three datasets. On the Douban-Book dataset, DBA-SR achieves scores of 0.1271, 0.1828, and 0.1744 on Recall@10, Recall@20, and NDCG@20, respectively, representing an improvement of approximately 0.5% to 1.3% compared to the best-performing baseline model, DVGRL. On the Douban-Movie dataset, the performance of this invention is improved to 0.0639, 0.1099, and 0.2659, respectively, with the largest gain (+2.0%) on the Recall@20 metric, demonstrating a significant advantage over XSimGCL and DVGRL. On the Yelp dataset, where user-item interactions are sparser, DBA-SR achieved scores of 0.0813, 0.1257, and 0.0766, respectively, which is an improvement of about 1.0% to 2.0% compared to the suboptimal model DcRec, fully demonstrating its robustness in scenarios with limited interaction data.

[0093] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules, units, or units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units, modules, or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0094] The units may or may not be physically separate. The components shown as units can be one or more physical units, meaning they can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0095] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0096] In particular, according to embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this invention. It should be noted that the computer-readable medium described above in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0098] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A social recommendation method based on denoising and preference balancing, characterized in that, Includes the following steps: Building a social graph; Denoise the social graph to obtain the denoised social graph; Using the normalized user item interaction matrix and the denoised social graph as input, probabilistic user embeddings are learned through two-branch variational encoding, and then user fusion representations are obtained through variational regularization, gating denoising and attention fusion. Based on user fusion representation, the interaction graph and social graph are jointly reconstructed, and the training target is obtained by optimizing the objective through preference balancing.

2. The social recommendation method based on denoising and preference balancing according to claim 1, characterized in that, Denoising a social graph to obtain a denoised social graph involves the following steps: Calculate the edge similarity of the social graph to obtain a similarity score; The social graph is denoised based on the similarity score to obtain the denoised social graph.

3. The social recommendation method based on denoising and preference balancing according to claim 2, characterized in that, The formula for calculating the similarity score is: in: Represents the similarity score. and Representing users respectively and A collection of interactive items, where deg(v) represents a node. The degree, It is a constant; Based on the similarity score, the social graph is denoised to obtain the denoised social graph. The specific steps include: Based on the set of similarity scores for all edges Calculate the distribution sensitivity threshold based on the mean and standard deviation: in: Indicates the distribution-sensitive threshold. and Let represent the mean and standard deviation of the similarity scores, respectively. Indicates the intensity of pruning; Introduce a quantile threshold to measure the sparsity of the overall edge distribution: in: Represents the sparsity threshold. Indicates the quantile threshold; The maximum value of the distribution sensitivity threshold and the sparsity threshold is used as the unified pruning threshold: When the edge weight score If the value is greater than θ, the edge is retained; otherwise, it is removed, thereby obtaining the denoised adjacency matrix and the denoised social graph. Symmetrical normalization is performed on the denoised social graph: in: This represents a social graph after noise reduction.

4. The social recommendation method based on denoising and preference balancing according to claim 3, characterized in that, Using a normalized user item interaction matrix and a denoised social graph as input, probabilistic user embeddings are learned through two-branch variational encoding. Then, through variational regularization, gated denoising, and attention fusion, a user fusion representation is obtained, which includes the following steps: Normalized user project interaction matrix With denoised social graph As input, probabilistic user embedding representations are extracted using two parallel variational encoders; these user embedding representations include social embeddings. and interaction representation ; Kullback–Leibler divergence is introduced as a variational regularization term to calculate the distance between the posterior distributions of behavioral and social branches and the standard Gaussian prior. Introducing content-based gating mechanisms for social embedding Perform filtering, combined with interactive representation Generate a gating vector, and obtain the gating social representation based on the gating vector; An attention fusion module is introduced to adaptively balance the importance of behavioral embedding and social embedding, and user fusion representation is obtained based on the gated social representation.

5. The social recommendation method based on denoising and preference balancing according to claim 4, characterized in that, Normalized user project interaction matrix With denoised social graph As input, probabilistic user embedding representations are extracted using two parallel variational encoders, specifically including the following steps: In each branch, the initial adjacency matrix of the interactive branches is set to Social branches are set as follows And the node representation is updated layer by layer through L-layer nonlinear propagation: The final output high-level features and The parameters are used to generate the Gaussian distribution, thus obtaining a probabilistic user representation: in, and These represent the mean vector and standard deviation vector of the latent representation of users from the perspective of interaction behavior, respectively. and Let these be the mean vector and standard deviation vector of the latent representation of a user from the perspective of social relationships, respectively. The distribution is then sampled using a reparameterization technique: in, and A random vector that follows a standard normal distribution is used to introduce randomness to characterize the uncertainty of the user's potential representation; the symbol ⊙ represents element-wise multiplication.

6. The social recommendation method based on denoising and preference balancing according to claim 5, characterized in that, The formula for calculating the distance between the posterior distributions of the behavioral and social branches and the standard Gaussian prior is as follows: In the formula: This represents the distance between the posterior distributions of the behavioral and social branches and the standard Gaussian prior. The formula for calculating the gate vector is: in, and Let represent the trainable weight matrices used for linear mapping of user interaction embeddings and social embeddings, respectively. This represents the corresponding bias vector; This represents a non-linear activation function used to restrict the values ​​of the gating vector within a preset range, thereby achieving soft selection of social information; The formula for calculating the gated social representation is as follows: Among them, symbols This indicates element-wise multiplication. This represents the trainable weight matrix used for feature enhancement of social embeddings. This represents the corresponding bias vector; the hyperbolic tangent function is used to introduce nonlinear mapping capabilities to enhance the expressive power of social features. Attention weights are concatenated from behavioral and social features: in, This represents a trainable weight matrix used to perform feature transformation on the concatenated joint features, which is used to capture the correlation between user behavior features and social features; This represents a trainable weight vector used to map the transformed features to attention scores; This represents a non-linear activation function used to normalize or range-constrain attention scores; (symbol) This represents a vector concatenation operation; The user fusion representation is as follows: in, This represents the user fusion representation.

7. The social recommendation method based on denoising and preference balancing according to claim 6, characterized in that, Based on user fusion representation, joint reconstruction of interaction graph and social graph is performed, and the training objective is obtained by optimizing the objective through preference balancing. The specific steps include: Reconstruct the interaction and social relationships to obtain social intimacy; Based on a dual-objective optimization mechanism, we obtain long-tail preference enhancement and popularity suppression; By constraining the consistency between the predicted social intimacy and the denoised social relationship through cross-entropy loss, a consistency constraint term is obtained; The training objective is obtained based on the long-tail preference enhancement, popularity suppression, and consistency constraint terms.

8. The social recommendation method based on denoising and preference balancing according to claim 7, characterized in that, Reconstructing interaction and social relationships to obtain social intimacy involves the following steps: Leveraging user embedding Embedded with items Calculate interaction preference scores to predict the strength of the association between users and items: in, This represents the item embedding representation used for reconstructing interaction relationships, which characterizes the latent semantic features of items in the prediction space; Using the same user embedding, social intimacy is calculated via a multilayer perceptron to recover the latent structural relationships in the denoised social graph: in, This represents the predicted social intimacy between user u and user v, used to characterize the strength of their social association in the latent feature space; and These represent the user embedding representations corresponding to user u and user v, respectively, and the user embedding representations serve as the input to the social decoding module; The predicted social relationship matrix is ​​obtained by summing up the predicted social intimacy values ​​of all users for (u,v).

9. The social recommendation method based on denoising and preference balancing according to claim 8, characterized in that, Based on a dual-objective optimization mechanism, long-tail preference enhancement and popularity suppression are achieved, specifically including the following steps: Increase the learning signal for rare items through a reverse popularity weighting mechanism: in, This represents the popularity of item i, used to characterize the frequency of the item's appearance in historical interaction data; This represents the long-tail enhancement coefficient, used to adjust the enhancement strength of the anti-prevalence weighting mechanism. Its value controls the extent to which the weight of low-prevalence items is increased. Subsequently, the reweighted binary cross-entropy loss function is defined as follows: Introducing a popularity regularization term suppresses overfitting of high-popularity samples and promotes a more balanced recommendation distribution: Where λ>1 is used to control the steepness of the penalty term; Interactive reconstruction targets integrate two types of mechanisms: Among them, coefficient The balance between enhancement and inhibition is controlled; The formula for calculating the consistency constraint is as follows: The formula for calculating the training objective is: Where β represents the weight coefficient of the KL divergence regularization term, which is used to adjust the influence of probability distribution constraints on the overall training objective.

10. A social recommendation system based on denoising and preference balancing, used to implement the social recommendation method based on denoising and preference balancing as described in any one of claims 1-8, characterized in that, include: The social graph building module is used to build social graphs; The structure-aware social graph pruning module is used to denoise the social graph and obtain a denoised social graph. The dual variational coding and social representation fusion module is used to learn probabilistic user embeddings through dual-branch variational coding, and then obtain the user fusion representation through variational regularization, gating denoising and attention fusion. The dual-graph decoding and preference balance optimization module is used to jointly reconstruct the interaction graph and social graph based on the user fusion representation, and obtain the training target through the preference balance optimization objective.

Images