A social denoising recommendation method based on side confidence invariant learning

Abstract

This invention discloses a social denoising recommendation method based on edge confidence invariant learning, belonging to the recommender system field within machine learning. The invention constructs an edge confidence estimation mechanism that integrates preference similarity and structural priors, and based on this, performs soft pruning and confidence-weighted propagation on the social graph to suppress noise diffusion from low-quality social edges at the source. Addressing social structural perturbations and noise distribution changes, the invention introduces a multi-environment generation mechanism, constructing multiple differentiated social environments under the guidance of edge confidence priors, and learning stable and effective preference signals through cross-environment invariance constraints, thereby reducing the model's dependence on environment-specific spurious correlations. Simultaneously, the invention designs a cross-domain gating fusion mechanism to selectively forget and enhance social representations, and combines cross-domain alignment constraints to alleviate semantic shifts between the social and recommendation domains, improving the controllability of social information injection.

Description

Technical Field

[0001] This invention belongs to the field of recommender systems in machine learning, specifically involving a personalized recommendation method based on graph neural network social denoising, invariant learning, and information bottleneck. Background Technology

[0002] In the context of information overload and a severe oversupply of content, recommender systems have become a key infrastructure for alleviating information overload, improving user experience, and enhancing platform efficiency, and are widely used in e-commerce, social media, short videos, and local services. Traditional collaborative filtering methods mainly rely on user-item interaction data to characterize preferences. However, in real-world systems, interactions are often naturally sparse: many users generate only a few actions, and new items lack sufficient exposure and feedback, making it difficult for the model to learn reliable representations, thus limiting its performance in cold start and long-tail distribution. To alleviate these problems, social recommendation introduces social relationships between users as auxiliary signals. Utilizing the assumption that "socially similar individuals have similar preferences," it transforms users' social connections into preference propagation paths, thereby obtaining a more complete and stable user profile even when interactions are sparse.

[0003] However, social relationships do not always accurately reflect preference similarity. Real-world social networks contain numerous connections unrelated to or even contradictory to consumer interests: users may follow celebrities but not endorse their consumption choices, or establish connections with colleagues, relatives, or weak ties with vastly different interests. Furthermore, fake accounts, mutual following, outdated relationships, and platform-driven non-interest-based social interactions all introduce preference noise. These connections unrelated to the recommendation objective can be summarized as "social noise." In graph representation-based social recommendation frameworks, social noise not only locally interferes with individual connections but also spreads and amplifies during message passing and multi-hop aggregation: a noisy edge may inject irrelevant preference patterns into the user representation and affect a wider range of neighbors through multi-layer propagation, ultimately leading to representation contamination, ranking shifts, and decreased recommendation performance. Therefore, effectively identifying and suppressing social noise while fully utilizing social information to alleviate sparsity has become a core challenge that must be addressed in the implementation of social recommendations.

[0004] Previous work has proposed various solutions to address social noise, such as soft filtering of neighbors using edge weights / attention, enhancing robustness through representation constraints, or suppressing dependence on accidental correlated signals by constructing perturbed views. These methods have achieved positive progress, but several key shortcomings remain:

[0005] (1) Insufficient generalization robustness: Many denoising strategies mainly identify "effective social edges" under a single training distribution. Once the social structure or noise pattern changes, the model is prone to failure due to spurious correlations specific to the environment.

[0006] (2) Intra-domain denoising and inter-domain fusion are separated: filtering edges only within the social graph cannot guarantee that the social representation will not generate semantic conflicts when injecting the recommendation representation; conversely, emphasizing only fusion gating may also mask the diffusion effect of structural layer noise.

[0007] (3) Insufficient granularity in denoising: Treating all edges equally with perturbation / deletion can easily damage truly valuable strong relationships, while neglecting to focus on and differentiate the "most difficult-to-distinguish uncertain edges." Furthermore, the credibility of social edges themselves often exhibits strong heterogeneity: there are strong connections with highly consistent preferences, highly questionable weak connections, and a large number of fuzzy connections in between that require task-driven discrimination. The lack of utilization of this heterogeneity will result in denoising that is either too conservative or too aggressive.

[0008] (4) Residual redundant noise in the representation layer: Even if the edges are weakened, noise may still remain in the representation space in the form of statistical dependence or redundant dimensions, continuously interfering with the ranking and training stability. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes a social denoising recommendation algorithm, ECIR, based on edge confidence invariant learning. The goal is to achieve robust denoising and controllable fusion from structure to representation and from training to testing while "fully utilizing social information". Specifically, to address the significant heterogeneity in social edge credibility, this invention constructs an edge confidence estimation mechanism that integrates preference similarity and structural priors. Based on this, soft pruning and confidence-weighted propagation are applied to the social graph to suppress noise diffusion from low-quality social edges at the source. To address social structural perturbations and noise distribution changes, this invention introduces a multi-environment generation mechanism. Guided by edge confidence priors, multiple differentiated social environments are constructed, and stable and effective preference signals are learned through cross-environment invariance constraints, thereby reducing the model's dependence on environment-specific spurious correlations. Simultaneously, this invention designs a cross-domain gating fusion mechanism to selectively forget and enhance social representations, and combines cross-domain alignment constraints to alleviate semantic shifts between the social and recommendation domains, improving the controllability of social information injection. Furthermore, this invention introduces HSIC-based information bottleneck constraints at the representation layer to further compress redundant dependencies in social enhancement representations, reducing latent noise residue and thus improving the model's generalization ability and training stability.

[0010] This invention discloses a social denoising recommendation method based on edge confidence invariant learning, which includes the following steps:

[0011] Step 1: Obtain user-item interaction data and social relationship data between users, construct user-item interaction graph and user social graph, and initialize vector representations for users and items. Use graph representation learning methods to obtain user representations in the recommendation domain and user representations in the social domain, respectively.

[0012] Step 2: For the quality differences of edges in the social graph, confidence modeling is performed on each social edge, and user preference similarity and node structure information are integrated to obtain edge confidence weights. The social graph is then weighted and propagated based on the edge confidence, thereby enhancing high-confidence edges and suppressing low-confidence edges during information propagation, achieving intra-domain social noise reduction.

[0013] Step 3: Based on the edge confidence, the social edge is divided into three segments to obtain the core edge, the noise edge and the uncertain edge. On this basis, multiple environment generators are constructed to perform probability sampling on the uncertain edge and generate multiple differentiated social subgraphs in different environments, thereby forming a multi-environment social structure view.

[0014] Step 4: Perform social representation learning and recommendation prediction in each environment, and introduce environment-level invariance constraints to keep the recommendation loss consistent in different environments, thereby learning a stable user preference representation across environments and reducing the model's dependence on perturbations of specific social structures.

[0015] Step 5: To address the semantic differences between the recommendation domain representation and the social domain representation, a cross-domain gating fusion mechanism is constructed. The forgetting gate filters out noisy information in the social representation, the enhancement gate strengthens effective information consistent with the recommendation goal, and the social representations in multiple environments are aggregated to obtain a stable social enhancement representation.

[0016] Step 6: Integrate the social enhancement representation into the recommendation domain user representation in a controllable manner, and introduce cross-domain alignment constraints to ensure that the fused representation maintains consistency in the semantic space, thereby reducing the negative migration risk caused by cross-domain information injection;

[0017] Step 7: Introduce information bottleneck constraints in the representation layer, measure and minimize the statistical dependency between the social augmentation representation and the original social representation, compress redundant information, suppress latent noise residue, and thus improve the compactness and robustness of the representation.

[0018] Step 8: Based on the fused user representation and item representation, perform preference prediction, construct the recommendation loss function, and perform joint optimization by combining environment invariance constraints, cross-domain alignment constraints, and information bottleneck constraints to obtain the final recommendation model.

[0019] Furthermore, the specific method of step 1 is as follows:

[0020] First, a learnable low-dimensional vector is assigned as an initial representation for each user and item, namely: user initial vector. Item initial vector LightGCN is used for multi-layer information propagation to capture higher-order cooperative signals, and the following is recorded: For users The set of neighboring items on the interaction graph.

[0021] For items The set of neighboring users, the first Layer propagation is defined as:

[0022] ;

[0023] To integrate higher-order information from different layers, inter-layer aggregation is used to obtain the final representation of the recommendation domain:

[0024] ;

[0025] in For the number of propagation layers, For layer aggregation weights;

[0026] For social graphs, user representations are enhanced through the aggregation of neighboring users, and this is recorded as follows: For users The set of neighbors on a social graph, given the social graph and its edge weights. , No. Layer propagation is defined as:

[0027] ;

[0028] in For border rights;

[0029] Similarly, the social domain multi-layer representations are aggregated to obtain the social domain user representation:

[0030] ;

[0031] in For the number of layers of social domain propagation, This is the aggregate weight.

[0032] Furthermore, the specific method for step 2 is as follows:

[0033] Preference for similarity perspective:

[0034] Preference strength is obtained by using a similarity metric based on the Jaccard coefficient and normalizing it to neighbors. :

[0035] ;

[0036] in, This represents the Jaccard coefficient. Indicates user A collection of interactive items Indicates user and users The Jaccard coefficient, denoted by user The collection of interactive items is ;

[0037] Degree centrality perspective:

[0038] Constructing socio-psychological strength using the degree centrality of neighbor nodes. And normalize the neighbors:

[0039] ;

[0040] Finally, the edge confidence is obtained by fusion. Combining information from both preference and structure, the edge confidence is defined as:

[0041] ;

[0042] in A larger value indicates that the edge is more likely to be a homogeneous connection and more beneficial to recommendations; conversely, a smaller value indicates that it is more likely to be a noisy edge.

[0043] Side confidence guides the spread of social graph messages:

[0044] For each user Preserved Neighbor Set First, the edge confidence is normalized within the neighborhood to obtain the propagation coefficient. :

[0045] ;

[0046] in, Indicates the user's social relationship edge. To prevent extremely small constants with a denominator of zero, let Indicates user In the The social domain representation of the layer, under the weighted LightGCN framework, is the first... The message aggregation of the layer is written as:

[0047] .

[0048] Furthermore, the specific method for step 3 is as follows:

[0049] Step 3.1: Based on the prior edge confidence, introduce... An environment generator constructs a multi-environment social view. First, the edge is divided into three segments: based on edge confidence. The social edge is divided into: core edge ( High noise edge () ), uncertain edges ( Core edges are always retained, noisy edges are removed first, and uncertain edges are used for environmental differentiation modeling.

[0050] Step 3.2: Create an environment generator: For each environment Define the edge retention probability. :

[0051] ;

[0052] in This represents the activation function. and This represents the learnable parameters in environment k. This represents the perturbation term under the k-condition;

[0053] , This represents the disturbance term under environment k. This represents the fully connected layer function, and the final sampling strategy is:

[0054] ;

[0055] Represents the core edge, Indicates an uncertain edge. This represents the Bernoulli function.

[0056] Further, step 4 involves performing social representation learning using a homeopathic function in each environment. for:

[0057] ;

[0058] This indicates how to calculate the recommendation loss under various environments. This represents the variance function.

[0059] Furthermore, the specific method for the cross-domain gating fusion mechanism in step 5 is as follows:

[0060] set up To recommend domain users, To represent users in the social domain under environment k, we first construct interaction items between the two domains: , Indicates element-wise multiplication;

[0061] (1) Forget gate: used to suppress noise or inconsistent information in the social domain.

[0062] ;

[0063] ;

[0064] in, Represents the forget gate function. This represents the activation function. For the Sigmoid function, For learnable parameter matrix, This represents the filtered user representation;

[0065] (2) Reinforcement gates: used to supplement and reinforce social information that is consistent with the recommendation goal.

[0066] ;

[0067] ;

[0068] in Indicates an enhanced door. This indicates the enhanced user response. It is the hyperbolic tangent function. These are learnable parameters.

[0069] The specific method for step 6 is as follows:

[0070] Step 6.1: Gated social representation within the environment;

[0071] By combining the outputs of forgetting and enhancement, we obtain the environment. Gated social expression :

[0072] ;

[0073] Step 6.2: Cross-environment aggregation;

[0074] use Each environment generates a social subgraph, and each environment receives a gated social representation. To obtain stable social enhancement representations, they are aggregated across environments to obtain aggregated representations. :

[0075] ;

[0076] in These are learnable parameters;

[0077] Step 6.3: Cross-domain integration and cross-domain alignment;

[0078] Social representation after cross-environment aggregation Then, it is injected into the user representation of the recommendation domain with a controllable strength to form the final user representation. :

[0079] ;

[0080] in These are learnable fusion parameters used to control the intensity of social information injection. For the recommendation field representation;

[0081] Step 6.4: Employ alignment loss based on cosine similarity:

[0082] ;

[0083] Represents the cross-domain loss function. Represents a set of users. This represents the cosine function.

[0084] Furthermore, the specific method for step 7 is as follows:

[0085] set up For primitive social representation, For social enhancement representation after cross-environment aggregation, the information bottleneck objective is to reduce [the impact of information bottlenecks] while ensuring predictive ability. and This reduces redundant information between elements, thereby suppressing reliance on the original social noise and making the representation more compact and task-relevant.

[0086] Dependency Suppression Based on HSIC: To achieve the above constraints, HSIC is used to measure the statistical dependency between the two sets of representations and minimize it; given sample representations HSIC is defined as: ; in, For a centered matrix, For the kernel matrix, Indicates the size of the data;

[0087] Pick , Define information bottleneck loss: .

[0088] Furthermore, the specific method for step 8 is as follows:

[0089] Using the inner product as the prediction function :

[0090] ;

[0091] For each environment k, a representation is obtained based on the social subgraph, and a prediction score is calculated. The recommended loss is:

[0092] ;

[0093] Representing the data set, the final objective function is:

[0094] ;

[0095] in The set of learnable parameters for the model. These are the weight hyperparameters for environment invariance, cross-domain alignment, information bottleneck enhancement, and L2 regularization, respectively.

[0096] Compared with existing technologies, the advantages of this invention are as follows:

[0097] 1. This invention introduces an edge confidence modeling mechanism, integrates user preference similarity and social structure information, quantitatively evaluates the reliability of social edges, and performs weighted propagation and soft pruning on the social graph accordingly. This suppresses the noise diffusion problem caused by low-quality social relationships from the source. Compared with recommendation methods that directly utilize the original social graph, this invention can significantly improve the accuracy and effectiveness of social information utilization.

[0098] 2. This invention constructs a multi-environment generation mechanism. By dividing social edges into layers and performing differentiated sampling in different environments, it simulates various possible social structure perturbation scenarios. Combined with environment invariance constraints, it learns cross-environment stable user preference representations, thereby effectively reducing the model's dependence on specific data distributions or noise patterns and improving its generalization ability and robustness in scenarios with changing distributions.

[0099] 3. This invention designs a cross-domain gating fusion mechanism, which uses forget gates and enhancement gates to fine-grainedly control social representations, thereby suppressing noise information and enhancing effective information. It also obtains stable social enhancement representations through cross-environment aggregation. At the same time, it introduces cross-domain alignment constraints to alleviate the semantic offset problem between the social domain and the recommendation domain, thereby reducing the risk of negative transfer and improving the controllability and expression quality of social information fusion.

[0100] 4. This invention introduces an information bottleneck mechanism based on statistical dependency constraints in the representation layer, which further eliminates the influence of implicit noise by compressing the redundant dependencies between the social augmentation representation and the original social representation. Attached Figure Description

[0101] Figure 1 This is a flowchart of the present invention.

[0102] Figure 2This is a diagram of the overall model architecture of the present invention. Detailed Implementation Plan

[0103] Key theoretical terms:

[0104] Graph Convolutional Networks (GCNs) achieve efficient learning of node feature representations by generalizing the application of traditional convolution operations to graph data. Its core lies in iteratively updating the feature representations of nodes using information from the node and its neighborhood. For a graph G=(V,E), the feature update of node v depends not only on its own attributes but also on the features of its neighboring nodes. This update mechanism is implemented through a specific formula.

[0105] ,

[0106] in, It is the first The node feature matrix of the layer It is the adjacency matrix A of graph G plus the self-connection (identity matrix) This allows each node to receive its own characteristic information. It is the corresponding degree matrix. It is the first The learnable weight matrix of the layer, and It is a non-linear activation function. GCN enables nodes to effectively capture topological information in the graph structure through this aggregation of local neighborhoods.

[0107] LightGCN demonstrates through ablation experiments that nonlinear activation and feature transformation structures in GCNs are not suitable for recommender systems. This is likely because each graph node in a recommender system only uses the user or item ID for model building and training; therefore, node information is not as rich as image information, and thus, a complex structure is unnecessary. The formula is as follows:

[0108]

[0109] in, It is the first The node feature matrix of the layer It is the adjacency matrix A of graph G plus the self-connection (identity matrix) ), It is the corresponding degree matrix.

[0110] Invariant learning: In practical applications, data often originates from different environments or distributions, such as differences in collection devices, time, location, or population groups. These differences typically lead to a distribution shift problem between training and test data. Traditional empirical risk minimization methods tend to rely on spurious correlations in the training data, resulting in a significant performance degradation when the distribution changes. Invariant learning aims to alleviate this problem. Its core idea is to learn a prediction mechanism that remains stable under different environments, thereby improving the model's generalization ability.

[0111] Formally, let the input variable be... , tag as Environment variables are The goal of invariant learning is to find representation functions. This ensures that the condition distribution satisfies invariance under different environments, that is: in any two environments... and Below, given representation back, The conditional distribution remains consistent.

[0112] Information Bottleneck: The information bottleneck method originates from information theory, and its goal is to achieve a balance between information compression and task-related information retention during representation learning. Specifically, given input variables... , middle representation and target variable The information bottleneck method aims to learn a compact representation. This allows the system to fully represent task-related information while discarding as much redundant information as possible that is irrelevant to the task. The optimization objective can be expressed as: minimizing... With input Mutual information between them, while maximizing With the goal Mutual information between the two means compressing the input information while retaining the information most useful for the prediction task. This is typically controlled by a trade-off parameter to manage the balance between the two.

[0113] To make the objectives and innovations of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0114] The specific embodiments of the present invention are as follows:

[0115] 1. Graph representation learning and embedding modeling first assign a learnable low-dimensional vector as an initial representation for each user and item: LightGCN is employed for multi-layer information propagation to capture high-order cooperative signals. For users The set of neighboring items on the interaction graph.

[0116] For items The set of neighboring users. Layer propagation is defined as:

[0117]

[0118] To integrate higher-order information from different layers, inter-layer aggregation is used to obtain the final representation of the recommendation domain:

[0119]

[0120] in For the number of propagation layers, For layer aggregation weights, in the context of social graphs, user representations can be augmented with social semantics through aggregation of neighboring users. Let... For users The set of neighbors on the social graph. Considering the differences in the quality of social edges, we will introduce edge confidence and multi-environment mechanisms to denoise social edges. Here, we first present a unified propagation form. For a given social graph (or a social subgraph in a certain environment) and its edge weights... , No. Layer propagation is defined as:

[0121]

[0122] in Let the edge weights be denoised as follows. When denoising is not considered, we can set... In social denoising recommendation algorithms based on edge confidence invariant learning, The outcome will be jointly determined by side confidence and environmental sampling, thereby suppressing the propagation impact of low-quality social connections. Similarly, the aggregation of multi-layer representations of the social domain yields the social domain user representation:

[0123]

[0124] in For the number of layers of social domain propagation, This is the aggregate weight.

[0125] 2. Edge confidence modeling and intra-domain social denoising: The first step is edge confidence modeling, which integrates preference similarity and degree centrality. From a preference similarity perspective: This invention uses a similarity metric based on the Jaccard coefficient and normalizes it to neighbors to obtain preference strength. :

[0126]

[0127] Among them, users are recorded. The collection of interactive items is Degree centrality perspective: This invention uses the degree centrality of neighboring nodes to construct socio-psychological strength. And normalize the neighbors:

[0128]

[0129] Finally, the edge confidence is obtained by fusion. Combining information from both preference and structure, the edge confidence is defined as:

[0130]

[0131] in A larger edge confidence value indicates that the edge is more likely to be a homogeneous connection and more beneficial to recommendations; conversely, a smaller value indicates that it is more likely to be a noisy edge. Edge confidence guides message propagation in social graphs: for each user... Preserved Neighbor Set First, the edge confidence is normalized within the neighborhood to obtain the propagation coefficient:

[0132]

[0133] in To prevent extremely small constants with a denominator of zero, let... Indicates user In the The social domain representation of the layer, under the weighted LightGCN framework, is the first... The message aggregation of the layer is written as:

[0134]

[0135] This allows high-confidence edges to have both a higher retention probability and a stronger propagation weight, while low-confidence edges are retained even after soft pruning, thus creating an in-domain denoising effect at the propagation layer.

[0136] 3. Multi-environment generation and environment-level invariant modeling: To improve the robustness of the model under different noise distributions and social perturbations, this invention introduces, based on edge confidence priors, […]. An environment generator constructs a multi-environment social view. This is to enable the model to learn a preference signal that is stable under various noise conditions. The first step is a three-segment partitioning: based on edge confidence. The social edge is divided into: core edge ( ): A high-confidence, stable, and reliable preference propagation channel that remains intact in all environments, avoiding the accidental rejection of valid social signals. High-noise side ( Low-confidence, questionable edges are handled by the environment generator for more significant perturbation or free pruning to maximize noise diversity. Uncertain edges ( Edges in the middle, the most difficult to distinguish, are the focus of environment differentiation modeling. Multi-environment learning helps the model filter out truly stable and effective edges. Core edges are always retained, noisy edges are removed first, and uncertain edges are used for environment differentiation modeling. Secondly, there is the environment generator: for each environment... Define the edge retention probability:

[0137]

[0138] in The final sampling strategy is as follows:

[0139] ,

[0140] Finally, there's environment-invariant learning: calculating the recommendation loss under various environments. And introduce invariance constraints:

[0141]

[0142] This encourages consistency across environments and helps learn stable preference signals.

[0143] 4. Cross-domain gating mechanism and cross-domain alignment: This invention introduces a cross-domain gating mechanism in each environment k to selectively forget and selectively enhance the social representation, and further reduces the difference between the two domain representations through cross-domain alignment constraints, so that the injected information is more focused on the part consistent with the recommendation goal.

[0144] Cross-domain gating: setting To recommend domain users, Let's represent users in the social domain under environment k. First, we construct the interaction items between the two domains: .

[0145] (1) Forget gate: used to suppress noise or inconsistent information in the social domain.

[0146]

[0147]

[0148] in For the Sigmoid function, This is the learnable parameter matrix.

[0149] (2) Reinforcement gates: used to supplement and reinforce social information that is consistent with the recommendation goal.

[0150]

[0151]

[0152] in It is the hyperbolic tangent function. These are learnable parameters.

[0153] In-environment gating social representation: Combining forgetting and enhancement outputs, the environment is obtained. The following gated social interaction indicates:

[0154]

[0155] The intuitive meaning of this structure is: Ensure that noise dimensions are suppressed. This ensures that beneficial signals are supplemented and amplified, thereby achieving a balance between "denoising and gain" during cross-domain injection.

[0156] Cross-environment aggregation: Because this invention adopts Each environment generates a social subgraph, and each environment receives a gated social representation. To obtain stable socially enhanced representations, they are aggregated across environments:

[0157]

[0158] in These are learnable parameters.

[0159] Cross-domain integration and cross-domain alignment: Social representations obtained after cross-environment aggregation Then, it is injected into the recommendation domain user representation with a controllable strength to form the final user representation:

[0160]

[0161] in These are learnable fusion parameters used to control the strength of social information injection. Furthermore, to further reduce the negative transfer risk caused by differences in representations between the two domains, a cross-domain alignment constraint is introduced to optimize the recommendation domain representation. With social enhancement Maintaining consistency in the semantic space. This invention employs an alignment loss based on cosine similarity:

[0162]

[0163] This constraint encourages the social injection portion to align with the semantic direction of the recommendation domain, thereby making the enhanced information obtained from cross-domain gating closer to the recommendation goal and further reducing the cross-domain transfer of noise from the social domain.

[0164] 5. Increased Information Bottleneck

[0165] Although social noise has been explicitly suppressed through edge confidence modeling, multi-environment learning, and cross-domain gating mechanisms, it may still remain implicitly in the representation space during graph propagation and cross-domain fusion. Examples include weakly correlated statistical co-occurrence, redundant structures introduced by higher-order propagation, and redundant dimensions from cross-domain fusion. While this type of representation layer noise does not directly correspond to specific edges, it continuously affects model learning and reduces generalization performance when distribution changes. To further purify the social enhancement representation, this invention introduces an information bottleneck mechanism in the cross-environment aggregation results. By retaining only the minimum sufficient information relevant to the recommendation goal and suppressing redundant dependencies with the original social representation, it reduces the risk of implicit noise transfer and improves model robustness.

[0166] Information bottleneck concept: For primitive social representation, This represents a socially enhanced representation after cross-environment aggregation. The information bottleneck objective is to reduce [the number of instances] while ensuring predictive ability. and This reduces redundant information between elements, thereby suppressing reliance on the original social noise and making the representation more compact and task-relevant.

[0167] Dependency Suppression Based on HSIC: To achieve the above constraints, this paper uses HSIC to measure the statistical dependency between the two sets of representations and minimizes it. Given sample representations HSIC is defined as: in For a centered matrix, It is the kernel matrix.

[0168] In this invention, take , Define information bottleneck loss: This constraint reduces the dependence of the social reinforcement representation on the original social representation, suppresses noise migration, and reinforces stable social information that is relevant to the recommendation task.

[0169] 6. Multi-task optimization and prediction layer: First, the inner product is used as the prediction function:

[0170]

[0171] For each environment k, a representation is obtained based on the social subgraph, and a prediction score is calculated. The recommended loss is:

[0172]

[0173] The final objective function is:

[0174]

[0175] in The set of learnable parameters for the model. These are the weight hyperparameters for environment invariance, cross-domain alignment, information bottleneck enhancement, and L2 regularization, respectively. The algorithm diagram is shown below. Figure 2 As shown.

[0176] 7. Using datasets to evaluate the model. After data preprocessing, this invention divides the dataset into training, validation, and test sets in an 8:1:1 ratio. To verify the effectiveness of this invention, three commonly used public datasets are selected: Douban, Epinions, and Yelp2018. Douban, from the Douban platform, contains user ratings of content (such as movies) and social relationships between users. This dataset contains 12,910 users, 22,291 items, and 591,296 interaction records, along with 169,150 social relationships. Its rating density is 0.21%, and its relationship density is 0.10%, showing a relatively balanced distribution of interaction and social sparsity, making it suitable for evaluating the model's performance in moderately sparse scenarios. Epinions, from an online product review community, contains user ratings and trust relationships. This dataset contains 22,160 users, 296,269 items, and 922,116 interaction records, along with 355,754 social relationships. Its rating density is only 0.01%, and its relationship density is 0.07%, indicating extremely sparse interactions. Furthermore, social relationships may contain connections unrelated to preferences, placing higher demands on the model's denoising capabilities. Yelp2018, from the Yelp platform, includes user interactions with merchants and friend relationships. This dataset contains 30,934 users, 22,123 items, and 441,304 interaction records, with 1,001,010 social relationships. Its rating density is 0.06%, and its relationship density is 0.10%. It boasts the largest social scale and most complex structure, making it suitable for evaluating the scalability and stability of models in large-scale social networks.

[0177] Experimental results show that the present invention improves the recommendation performance on public datasets in terms of key metrics such as Recall and NDCG.

[0178] The experimental environment and parameter settings are as follows: the relevant experiments are run on the server, and the environment configuration used is shown in Table 1.

[0179] Table 1 Experimental Environment Configuration Information

[0180]

[0181] The following cutting-edge recommendation models were selected as benchmarks for comparison:

[0182] HGSR: A social recommendation model based on hyperbolic graph learning. It embeds social structures in hyperbolic space and combines pre-trained social embeddings with heterogeneous graph learning to cope with network heterogeneity and reduce the impact of diffusion noise.

[0183] GDMSR: A denoising and enhanced social recommendation framework that uses preference-guided social graph denoising to preserve informative social relationships, and combines self-correcting curriculum learning with adaptive denoising strategies to mitigate the negative impact of noisy social connections.

[0184] SHaRe: Improves the quality of social graphs and enhances recommendation performance at the structural level by rewiring highly homogeneous social relationships and removing low homogeneous connections.

[0185] GBSR: A graph denoising and bottleneck learning framework that includes preference-guided social graph refinement and bottleneck learning to learn the "minimum but sufficient" denoised social structure, thereby improving model generalization and robustness.

[0186] CLSRec: A social recommendation method that reconstructs a smoother / denoising social view by performing low-rank SVD on the interaction matrix. This reconstructed view is then compared and aligned with the original social graph view to suppress social noise propagation and enhance user representation. Simultaneously, it combines distillation and gating fusion to coordinate social and behavioral interests, employing BPR and contrastive loss for joint training.

[0187] The following two widely used evaluation metrics are adopted:

[0188] Recall@K is a recommendation system evaluation metric, which measures the percentage of the top K items recommended to a user that match items the user is genuinely interested in.

[0189] NDCG@K considers the ranking position of items in the recommendation list, assigning higher weights to correct recommendations that appear earlier, and comprehensively evaluates the model's accuracy and ranking ability.

[0190] As shown in Table 2, the experimental results demonstrate that the present invention achieves significant performance improvements on all datasets. Specifically, on the Douban dataset, R@20 and N@20 reach 0.1686 and 0.1182, respectively; on the Epinions dataset, R@20 and N@20 reach 0.0605 and 0.0359, respectively; and on the Yelp2018 dataset, R@20 and N@20 reach 0.1228 and 0.0669, respectively. It can be seen that the present invention provides a certain improvement over previous models.

[0191] Table 2. Recommendation results of this invention on three public datasets.

[0192]

Claims

1. A social denoising recommendation method based on edge confidence invariant learning, the method comprising the following steps: Step 1: Obtain user-item interaction data and social relationship data between users, construct user-item interaction graph and user social graph, and initialize vector representations for users and items. Use graph representation learning methods to obtain user representations in the recommendation domain and user representations in the social domain, respectively. Step 2: For the quality differences of edges in the social graph, confidence modeling is performed on each social edge, and user preference similarity and node structure information are integrated to obtain edge confidence weights. The social graph is then weighted and propagated based on the edge confidence, thereby enhancing high-confidence edges and suppressing low-confidence edges during information propagation, achieving intra-domain social noise reduction. Step 3: Based on the edge confidence, the social edge is divided into three segments to obtain the core edge, the noise edge and the uncertain edge. On this basis, multiple environment generators are constructed to perform probability sampling on the uncertain edge and generate multiple differentiated social subgraphs in different environments, thereby forming a multi-environment social structure view. Step 4: Perform social representation learning and recommendation prediction in each environment, and introduce environment-level invariance constraints to keep the recommendation loss consistent in different environments, thereby learning a stable user preference representation across environments and reducing the model's dependence on perturbations of specific social structures. Step 5: To address the semantic differences between the recommendation domain representation and the social domain representation, a cross-domain gating fusion mechanism is constructed. The forgetting gate filters out noisy information in the social representation, the enhancement gate strengthens effective information consistent with the recommendation goal, and the social representations in multiple environments are aggregated to obtain a stable social enhancement representation. Step 6: Integrate the social enhancement representation into the recommendation domain user representation in a controllable manner, and introduce cross-domain alignment constraints to ensure that the fused representation maintains consistency in the semantic space, thereby reducing the negative migration risk caused by cross-domain information injection; Step 7: Introduce information bottleneck constraints in the representation layer, measure and minimize the statistical dependency between the social augmentation representation and the original social representation, compress redundant information, suppress latent noise residue, and thus improve the compactness and robustness of the representation. Step 8: Based on the fused user representation and item representation, perform preference prediction, construct the recommendation loss function, and perform joint optimization by combining environment invariance constraints, cross-domain alignment constraints, and information bottleneck constraints to obtain the final recommendation model.

2. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method for step 1 is as follows: First, a learnable low-dimensional vector is assigned as an initial representation for each user and item, namely: user initial vector. Item initial vector LightGCN is used for multi-layer information propagation to capture higher-order cooperative signals, and the following is recorded: For users The set of neighboring items on the interaction graph. For items The set of neighboring users, the first Layer propagation is defined as: ； To integrate higher-order information from different layers, inter-layer aggregation is used to obtain the final representation of the recommendation domain: ； in For the number of propagation layers, For layer aggregation weights; For social graphs, user representations are enhanced through the aggregation of neighboring users, and this is recorded as follows: For users The set of neighbors on a social graph, given the social graph and its edge weights. , No. Layer propagation is defined as: ； in For border rights; Similarly, the social domain multi-layer representations are aggregated to obtain the social domain user representation: ； in For the number of layers of social domain propagation, This is the aggregate weight.

3. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method for step 2 is as follows: Preference for similarity perspective: Preference strength is obtained by using a similarity metric based on the Jaccard coefficient and normalizing it to neighbors. : ； in, Represents the Jaccard coefficient. Indicates user A collection of interactive items Indicates user and users The Jaccard coefficient, denoted by user The collection of interactive items is ; Degree centrality perspective: Constructing socio-psychological strength using the degree centrality of neighbor nodes. And normalize the neighbors: ； Finally, the edge confidence is obtained by fusion. Combining information from both preference and structure, the edge confidence is defined as: ； in A larger value indicates that the edge is more likely to be a homogeneous connection and more beneficial to recommendations; conversely, a smaller value indicates that it is more likely to be a noisy edge. Side confidence guides the spread of social graph messages: For each user Preserved Neighbor Set First, the edge confidence is normalized within the neighborhood to obtain the propagation coefficient. : ； in, Represents the user's social relationship edge. To prevent extremely small constants with a denominator of zero, let Indicates user In the The social domain representation of the layer, under the weighted LightGCN framework, is the first... The message aggregation of the layer is written as: 。 4. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method for step 3 is as follows: Step 3.1: Based on the prior edge confidence, introduce... An environment generator constructs a multi-environment social view. First, the edge is divided into three segments: based on edge confidence. The social edge is divided into: core edge ( High noise edge () ), uncertain edges ( Core edges are always retained, noisy edges are removed first, and uncertain edges are used for environmental differentiation modeling. Step 3.2: Create an environment generator: For each environment Define the edge retention probability. : ； in This represents the activation function. and This represents the learnable parameters in environment k. Represents the perturbation term under the k-condition; , This represents the perturbation term under environment k. This represents the fully connected layer function, and the final sampling strategy is: ； Represents the core edge, Indicates an uncertain edge. This represents the Bernoulli function.

5. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, Step 4 involves performing social representation learning using a follow-mode function in each environment. for: ； This indicates how to calculate the recommendation loss under various environments. This represents the variance function.

6. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method of the cross-domain gating fusion mechanism in step 5 is as follows: set up To recommend domain users, To represent users in the social domain under environment k, we first construct interaction items between the two domains: , Indicates element-wise multiplication; (1) Forget gate: used to suppress noise or inconsistent information in the social domain. ； ； in, Represents the forget gate function. This represents the activation function. For the Sigmoid function, For learnable parameter matrix, This represents the filtered user representation; (2) Reinforcement gates: used to supplement and reinforce social information that is consistent with the recommendation goal. ； ； in Indicates an enhanced door. This indicates the enhanced user response. It is the hyperbolic tangent function. These are learnable parameters.

7. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method for step 6 is as follows: Step 6.1: Gated social representation within the environment; By combining the outputs of forgetting and enhancement, we obtain the environment. Gated social expression : ； Step 6.2: Cross-environment aggregation; use Each environment generates a social subgraph, and each environment receives a gated social representation. To obtain stable social enhancement representations, they are aggregated across environments to obtain aggregated representations. : ； in These are learnable parameters; Step 6.3: Cross-domain integration and cross-domain alignment; Social representation after cross-environment aggregation Then, it is injected into the user representation of the recommendation domain with a controllable strength to form the final user representation. : ； in These are learnable fusion parameters used to control the intensity of social information injection. For the recommendation field representation; Step 6.4: Employ alignment loss based on cosine similarity: ； Represents the cross-domain loss function. Represents a set of users. This represents the cosine function.

8. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method for step 7 is as follows: set up For primitive social representation, For social enhancement representation after cross-environment aggregation, the information bottleneck objective is to reduce [the impact of information bottlenecks] while ensuring predictive ability. and This reduces redundant information between elements, thereby suppressing reliance on the original social noise and making the representation more compact and task-relevant. Dependency Suppression Based on HSIC: To achieve the above constraints, HSIC is used to measure the statistical dependency between the two sets of representations and minimize it; given sample representations HSIC is defined as: ; in, For a centered matrix, For the kernel matrix, Indicates the size of the data; Pick , Define information bottleneck loss: .

9. The social denoising recommendation method based on edge confidence invariant learning as described in claim 1, characterized in that, The specific method for step 8 is as follows: Using the inner product as the prediction function : ； For each environment k, a representation is obtained based on the social subgraph, and a prediction score is calculated. The recommended loss is: ； Representing the data set, the final objective function is: ； in The set of learnable parameters for the model. These are the weight hyperparameters for environment invariance, cross-domain alignment, information bottleneck enhancement, and L2 regularization, respectively.

Images