Social recommendation denoising method based on ternary resonance and polarity modulation graph neural network

Abstract

The invention discloses a social recommendation denoising method based on a ternary resonance principle, bipolar perception and a polar modulation graph neural network. The method comprises the following steps: firstly, according to a ternary resonance principle, performing prior auditing on a social link by analyzing a user-article ternary interaction structure, and generating a topological mask to identify a high-confidence link; secondly, refining the original social graph by using the mask, and injecting the purified relationship as a structured prior into a user-article interaction graph for heterogeneous fusion; then, on the enhanced graph, through a bipolar perception and modulation graph neural network, utilizing a polarization modulation ordinal number to dynamically regulate and control information flow, and learning node characterization capable of distinguishing whether a user is good or not; and finally, constructing a composite objective function containing polarization ordinal number constraint and Bayesian personalized ordering, and optimizing the model by minimizing the function. Social noise can be effectively filtered, user positive and negative feedback is deeply modeled, and recommendation accuracy is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a social recommendation system using bipolar perception representation learning, ternary resonance principle, polar modulation graph neural network, Bayesian personalized ranking, polarization ordinal constraint, graph structure denoising and enhancement. BACKGROUND

[0002] As a key research direction in the field of information retrieval and data mining, in recent years, with the vigorous development of e-commerce, online streaming media, local life service and short video platform, etc. Application scenarios have shown important academic research and industrial application value. Its core task is to alleviate the problem of information overload, by modeling user historical behavior data and item attributes, predicting the user's potential interest, so as to realize personalized information distribution. Researchers have proposed a series of classic models such as DeepFM based on feature cross, YouTube DNN based on deep neural network, and DIN fusing sequence attention, etc. These models greatly improve user experience and business conversion efficiency in their respective business scenarios, and also mark the evolution of recommendation technology from traditional naive collaborative filtering and matrix decomposition to deep learning architecture that can capture deep and nonlinear relationships between users and items.

[0003] However, data sparsity, cold start and long tail effect are always inherent challenges faced by recommendation systems. These problems seriously restrict the generalization ability and recommendation accuracy of the model, limit the accuracy of the model learning user preference representation, especially for new users or cold items, it is difficult to provide personalized services. To cope with these challenges, the introduction of heterogeneous auxiliary information has become the mainstream paradigm, such as social networks, knowledge graphs and multi-modal information, aiming to build a multi-view representation space for users and items, to make up for the lack of main interaction data through the complementarity between different information sources, to enrich the semantic information of users and items through external knowledge, and to enhance the representation learning of users and items. This evolution of methodology marks the paradigm shift of recommendation systems from relying on single collaborative signal to building hybrid information network models. Among them, social networks explicitly depict the trust or social connection between users, which has the natural advantage of revealing the homogeneity of user preferences, that is, users connected to each other tend to have similar interests, and this homogeneity assumption provides a valuable transfer path for the model to reason about preferences under sparse interaction data. In addition, compared with knowledge graphs that require complex knowledge extraction and alignment, or multi-modal information that relies on large-scale pre-trained models, social networks have the advantage of relatively lightweight data structure and direct use without external knowledge, so they have become a key auxiliary information source in social recommendation systems and have been widely applied and researched.

[0004] Social recommendation systems aim to leverage the topology of social networks to enhance user representation learning and thus improve recommendation performance. In the framework of graph neural networks, social networks are usually treated as a user homogenous graph, which is jointly learned with a user-item heterogeneous bipartite graph. The preference information of social neighbors is aggregated onto the center user node through a message passing mechanism. However, the semantic gap between social connections and user preference homogeneity leads to inherent and widespread noise in real-world social networks. Specifically, social links (e.g., colleagues, alumni relationships) do not always equate to consistency in interest preferences, and users' tastes can even be completely opposite. Such preference-inconsistent social links can seriously mislead the modeling process of user interests and introduce false inductive bias. During the information propagation process of graph neural networks, these noisy links can cause distortion and diffusion of preference information, even exacerbate the "over-smoothing" problem, making the representations of users in the social circle tend to homogenization and lose individualization. Therefore, existing researches are committed to solving the denoising problem in social networks. The current mainstream denoising methods mostly adopt strategies based on node representation similarity or self-supervised learning signals. For example, some methods calculate the cosine similarity of user embedding vectors to determine the reliability of social relationships, or use a contrastive learning framework to distinguish strong and weak social relationships. Such methods are usually classified as post-hoc or implicit denoising strategies, because they rely on the intermediate products generated during the model learning process (i.e., node embeddings) to make judgments. However, such methods are highly dependent on the quality of node embedding representation, and are prone to the dilemma of circular argumentation: noisy social relationships pollute the learning of node embeddings, and the contaminated embeddings in turn lead to the failure of the denoising task. When the original graph noise is serious, it is easy to fall into a vicious cycle of "noise input-noise output". How to design a more robust priori denoising paradigm that does not completely rely on the quality of embeddings is a key challenge.

[0005] In addition, the ultimate goal of social network denoising is to serve the recommendation task, so its process should be guided by the refinement of user-item interaction information, which means that the denoising mechanism itself needs to be able to "understand" the user's preferences, rather than just analyzing the topology of the graph. However, the existing methods are rough in the granularity of mining interaction information. Specifically, most methods only use positive polarity interaction information (such as clicks, purchases) of user-item, while completely ignoring negative polarity interaction information (such as low ratings, not interested) which contains rich preference signals. Negative polarity provides strong signals about the boundaries of user preferences, and is the key to distinguishing different user groups, and even different interest dimensions of the same user. From the perspective of decision boundary, negative polarity provides clear sample signals, which is crucial for learning more discriminative user preference boundaries in latent space. Compared with implicit negative samples inferred by the model from unobserved interactions, explicit negative feedback generated by users is a more valuable supervision signal. This one-sided use of interaction information makes the model unable to obtain the key clues for judging the reliability of social relationships from the "conflict" of user opinions, and the model loses the ability to infer the preference heterogeneity of two users by observing whether they hold opposite opinions on the same item, leading to unsatisfactory results of social network denoising, and even possibly incorrectly pruning some weak but critical preference consistency links, introducing new information bias.

[0006] In order to accurately and efficiently deal with the above two challenges, the present application proposes a social recommendation denoising method based on ternary resonance and polar modulation graph neural network. First, a social relationship auditing mechanism based on ternary resonance principle is proposed. The mechanism follows the social ternary resonance principle to make a priori judgment on the internal consistency of social links before feature learning. By deconstructing the triangular structure formed by users interacting with common items, the consistency of preferences behind each social relationship is quantified. This prior analysis at the graph topology level rather than the feature space level can fundamentally reduce the dependence on node embedding quality. Second, a structure consistency driven polar heterogeneous graph enhancement strategy is proposed. Based on the aforementioned auditing mechanism, a topology mask matrix is generated, and the high confidence social relationships identified are used as structured prior to purify the original social network and remove noise links with conflicting preferences. Effectively integrating social topology structure and user behavior patterns, the method provides a more information-rich graph base for subsequent representation learning. Third, a node-level adaptive polarization modulation mechanism is proposed. The mechanism can dynamically calculate the polarization modulation coefficient of each node according to the local polar neighborhood statistical characteristics, to realize fine modeling of the polarity in the graph. Fourth, an adaptive bipolarity-aware high-order information aggregation graph encoder is proposed. The encoder can dynamically modulate the information flow strength from positive and negative polar feedback neighbors based on the aforementioned polarization modulation coefficient during the information propagation process of the polar graph neural network. This fine-grained multi-channel information flow control mechanism enables the model to capture more complex asymmetry in user preferences, thereby learning more robust node representations that can fully reflect user preferences. Fifth, a joint optimization framework based on polar ordinal constraint is constructed. The framework integrates polar ordinal constraint and pairwise learning paradigm to form a compound objective function, and trains the model end-to-end to ensure that the learned node representation can effectively distinguish positive and negative preferences of users and accurately sort personalized items. SUMMARY

[0007] To solve the problem of data sparsity and the significant semantic gap between social relationships and preferences (i.e., noise) in current social recommendation datasets (such as Douban-Book, Epinions, CiaoDVD, etc.), and the limitations of existing social recommendation methods in denoising mechanisms, which rely on posterior embedding quality and lead to insufficient robustness and inadequate mining of key discriminative signals contained in user negative feedback information, the present application proposes a social recommendation denoising method based on ternary resonance and polar modulation graph neural networks. The main content of the present application includes: quantifying the preference consistency between users through a social link prior evaluation algorithm based on ternary closed-loop structure, and generating a topology mask to perform graph structure purification and heterogeneous fusion; then, designing a bipolar perception information propagation network based on polarization modulation coefficient on the refined graph to learn the deep representation of users and items; finally, constructing a joint optimization framework based on pairwise ranking learning, optimizing the model by minimizing a composite loss function containing Bayesian personalized ranking loss and polarization ordinal constraint, and performing vector dot product prediction preference score to recall recommended items.

[0008] The main steps of the social recommendation denoising method based on ternary resonance and polar modulation graph neural networks proposed in this paper are as follows:

[0009] Step one, a user preference consistency quantification method based on common interaction history is proposed. This method first discretizes the explicit feedback signal of user-item interaction through a pre-set threshold, constructs a polar bipartite graph, and generates positive and negative polarity adjacency matrices according to the polarity of the edges corresponding to user preference and dislike for items. Subsequently, for any binary user group in the social network topology, this method analyzes the local network topology without relying on any learned embedding representation, especially the ternary closed-loop structure formed by the common interaction items, and follows the principle of structural ternary resonance in social networks to design a consistency score function. This function is used to quantify the degree of preference consistency and preference conflict exhibited by the user group in historical interaction behavior, thereby generating a prior confidence score for each social link based on structured evidence.

[0010] Step two, a social graph purification and multi-relation graph fusion strategy based on topology mask is proposed. Based on the quantitative results of the audit mechanism in step one, a binary topology mask matrix is generated, which can be regarded as a relationship filter acting on the social graph. By applying this mask to the original social graph, a relationship pruning operation is performed to filter out low-confidence noise links that are judged to be preference conflicts. This process generates a more refined social subgraph that better reflects the true preference homogeneity. Furthermore, this refined subgraph is used as a structured prior, and through linear superposition of adjacency matrices, it is fused with the polarity bipartite graph of user-item interactions across domains to ultimately construct a unified, multi-view enhanced heterogeneous polarized information network, providing a more information-rich and structurally reliable graph base for downstream representation learning tasks.

[0011] Step three, a node-level adaptive polarization modulation coefficient is proposed. This coefficient can effectively decouple and finely model the positive and negative polarity information in the graph. It is dynamically calculated based on the local polarity neighborhood statistical characteristics of each node, and its value reflects the relative dependence of the node on information flow from positive and negative polarity neighbors. The introduction of this coefficient provides a core modulation mechanism for adaptive weighted information aggregation in subsequent graph neural networks and is a prerequisite for fine-grained information flow control.

[0012] Step four, an information aggregation network for bipolar graphs is proposed. On this enhanced heterogeneous information network, a polarity-modulated graph neural network encoder is designed to learn the dense embedding vectors of users and items in the low-dimensional latent space. To effectively decouple and model the polarity information in the graph, the node-level polarization modulation coefficient calculated in step three is integrated into the neighborhood information aggregation function to form a dynamic information flow modulation mechanism for adaptively adjusting the aggregation weights from positive and negative polarity neighbors. Through nonlinear transformation and aggregation of multi-order neighborhood information, this encoder can learn node representations that comprehensively capture user preferences and are more discriminative and robust.

[0013] Step five, a joint optimization and recommendation recall framework based on polarization ordinal constraint is proposed. Using the final user and item representation vectors learned in step three, the relevance scores of users and items in the latent space are estimated through inner product operations. A bipolar-aware polarization ordinal constraint term is proposed to explicitly promote the two polarizations of positive samples, thereby providing a clear decision boundary. For model training, this polarization ordinal constraint will be combined with a pairwise learning paradigm loss term based on Bayesian personalized ranking (BPR) to form a joint loss function. Optimization algorithms based on stochastic gradient descent and its variants are used to minimize this composite objective function to iteratively update the model parameters. In the model inference stage, the personalized ranking scores of all candidate items for the target user are calculated, and a Top-K recommended item list is generated for recall.

[0014] Advantages:

[0015] 1. Significantly improves the accuracy and personalization level of recommendation results. The design scheme deeply integrates the purified social topology structure and the user's bidirectional interaction behavior through an innovative collaborative mechanism. The refined social graph provides high-quality structured paths for the propagation of preference information, while the polarization modulation graph encoder ensures the richness and discriminability of the propagated content. The synergy of the two makes the semantic alignment of users and items in the latent feature space more accurate, resulting in higher precision and novelty in the recommendation ranking task.

[0016] 2. Achieves high-fidelity semantic purification of social relationship graphs. The design scheme creatively introduces the social ternary resonance principle as a priori, structure-based auditing tool independent of embedding learning. It directly assesses the semantic consistency of social links by deconstructing the micro-ternary interaction patterns between users at the behavioral data level. This method surpasses traditional post-hoc denoising paradigms that rely on representation similarity, enabling high-fidelity purification of the topology structure of the original social network, making it more accurately reflect the homogeneity assumption between user groups, and providing a more reliable data foundation for all downstream tasks.

[0017] 3. Promotes the overall and multidimensional depth modeling of user preference profiles. The adaptive bipolar perception encoder in the design scheme is one of the first architectures that can decouple and adaptively integrate explicit polarity feedback signals in social recommendation. By dynamically balancing information flow from different polarity neighborhoods, the model can learn a more discriminative decision boundary, expanding user preferences from the traditional "like-unknown" binary space to the more refined "like-hate-unknown" ternary semantic space, thus constructing a more three-dimensional and complete user preference model.

[0018] 4. Enhances the information diffusion efficiency and generalization ability of the model in sparse data environments. The design scheme injects the purified social network as a reliable structural prior into the originally sparse user-item bipartite graph. This graph fusion strategy effectively constructs new high-order information propagation paths, greatly enhancing the diffusion efficiency and coverage of preference signals in the graph structure. Therefore, even in scenarios with extremely scarce interaction data such as cold-start users or long-tail items, the model can still learn meaningful representations through these high-quality "social shortcuts", thus demonstrating stronger generalization performance. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 . Method framework flowchart

[0020] Figure 2 . Social link purification and heterogeneous polar network enhancement algorithm flowchart

[0021] Figure 3 . Bipolar-aware polar modulation graph neural network model flowchart DETAILED DESCRIPTION

[0022] Step one, perform a priori preference consistency audit of social links. Without relying on any learned embedding representation, a priori preference consistency confidence is quantified for each link in the original social network by deconstructing the local topology of the user-item interaction graph. The overall process first polarizes the interaction data for preprocessing, and then extracts the preference relationship between user pairs in the three-element closed loop structure formed based on common interaction items through matrix operation, and finally generates a topology mask for the subsequent steps.

[0023] (1) First, the explicit rating feedback in the user-item interaction history is mapped by a preset threshold to construct a polar binary graph using the following formula:

[0024]

[0025] The edge set of the graph is divided into positive polarity subgraph and negative polarity subgraph corresponding to the "like" and "dislike" semantics of the user-item interaction, respectively. This process generates two sparse adjacency matrices, denoted as positive polarity adjacency matrix and negative polarity adjacency matrix, respectively.

[0026] (2) Based on the principle of social ternary resonance, the preference consistency between user pairs can be inferred from their relationship with common neighbors (i.e., items that have been interacted with). The ternary closed loop structure is extracted by using matrix multiplication to efficiently capture these two-hop range ternary structures. Among them, the path formed by two same polarity links between user pairs is considered as a "consistency" path; while the path formed by two different polarity links is considered as a "conflict" path.

[0027] (3) The extracted preference relationship paths are aggregated and binarized to generate a global topology mask. The consistency mask and the conflict mask are calculated as follows, which respectively identify which user pairs have at least one preference consistency or preference conflict behavior in the entire interaction history:

[0028] M + =B(A + ·A + +A-·A-)

[0029] M-=B(A + ·A-+A-·A + )

[0030] Step two, topology mask based heterogeneous polarized graph structure enhancement. Through the prior preference consistency evaluation in step one, a structured denoising operation is performed on the original social network, and the purified high-quality social relationship is taken as a kind of topology prior, which is cross-domain fused with the user-item interaction graph to build a more reliable graph base.

[0031] (1) To accurately identify the user pairs that only show consistency without any conflict, the ambiguous relationship that both signals exist is calculated first. Then, the part of ambiguous relationship is subtracted from the consistency mask to obtain a pure consistency mask. The mask accurately identifies those high-confidence user pairs that show preference consistency in all common interactions, and the calculation process is as follows:

[0032] M = M + - (M + ⊙M-)

[0033] (2) The pure consistency mask obtained in the last step is taken as a relationship filter and applied to the adjacency matrix of the original social network. Through element multiplication operation, a topological pruning is performed on the original social graph to remove all low-quality links that are not identified, thereby generating a semantically purified social relationship graph, and the adjacency matrix is calculated as follows:

[0034] S * = S ⊙ M

[0035] (3) After that, the purified social relationship graph is taken as a structured enhancement information and injected into the user-item interaction graph through matrix addition operation. This process fuses the social homogeneity information and user behavior information at the structure level to generate enhanced positive and negative polarity adjacency matrices, which together constitute a unified, multi-view polarized heterogeneous information network:

[0036]

[0037] Step three, based on the enhanced heterogeneous polarized information network obtained in step two, the adaptive polarized modulation coefficient is calculated to provide dynamic weight guidance for the subsequent bipolar message passing, so as to realize the fine modeling of the polarized information in the graph.

[0038] (1) First, the structure of the heterogeneous information network is investigated, and by traversing each node in the network, the degree polarity information of the local neighborhood is analyzed, and the statistical features in the neighborhood environment are extracted.

[0039] (2) After that, the unique polarization modulation coefficient of each node is dynamically calculated according to the acquired statistical characteristics of the neighborhood polarity of each node, which is used to quantify the degree of dependence on positive and negative polarity information in subsequent information aggregation. The calculation formula of the coefficient is as follows:

[0040]

[0041] Step four, bipolar perception graph convolutional representation learning. Based on the purified heterogeneous enhanced graph obtained in step two and the polarization modulation coefficient calculated in step three, a novel polarity modulation graph neural network encoder is involved, which performs multi-layer information propagation and aggregation on the enhanced heterogeneous information graph to learn a low-dimensional dense representation that can fully capture user preferences.

[0042] (1) At each layer of the graph neural network, the update of the node representation follows a bipolar perception message passing paradigm. First, the symmetric normalization is performed on the enhanced adjacency matrix. Then, the next layer representation of the node is composed of the aggregation information of its last layer representation and its polar neighborhood, where the information flow from different polar neighborhoods is dynamically regulated by the polarization modulation coefficient matrix, and its update rule is as follows:

[0043]

[0044] (2) To capture the neighborhood structure information of different distances and prevent over-smoothing, the output representation of all graph convolutional layers including the initial embedding layer is element-wise average-pooled. This operation fuses the neighborhood information of different orders to generate the final node representation rich in multi-scale structure information.

[0045]

[0046] Step five, based on the node representation learned in step four, construct the joint optimization objective and perform recommendation. To achieve the final optimization objective of the model, a composite loss function needs to be minimized to learn all trainable parameters, and after the model converges, the learned user and item representations are used to perform efficient personalized recommendation.

[0047] (1) First, define the relevance measure function of users and items in the latent semantic space. For any user u and item i, their personalized preference score is estimated by the inner product of their final representation vectors learned in step three:

[0048]

[0049] This operation can be regarded as projecting user preferences and item attributes into the same low-dimensional space and quantifying the matching degree through the geometric relationship between vectors.

[0050] (2) To further utilize the negative feedback signals in the interaction data, a bipolar-aware ordinal-constrained loss term is introduced. This loss term aims to explicitly impose constraints on the predicted scores of polarized links, i.e., requiring the scores of positive polarity interactions to be as high as possible, while the scores of negative polarity interactions to be as low as possible. This can help the model learn more discriminative decision boundaries. To further endow the model with ranking ability, the pairwise learning optimization paradigm is adopted. The core idea of BPR loss is to maximize the gap between the predicted scores of positive item i and non-interacted item j. This loss term is implemented through a log-sigmoid function, whose formula is as follows:

[0051]

[0052] The above loss term is linearly weighted with the L2 regularization term to form the final composite objective function of the model. The L2 regularization term is used to prevent model overfitting and improve generalization ability. The overall loss function is represented as follows:

[0053]

[0054] (3) In the inference phase, for any target user, the model will calculate the preference score of each item in the candidate item set. Then, all candidate items are ranked in descending order according to their predicted scores to generate a complete personalized ranking list, and K items are recalled for exposure in the following way:

[0055]

[0056] When the social recommendation denoising method based on ternary resonance and polarity modulation graph neural network is applied to the social recommendation task, not only can it effectively identify and filter noisy social relationships, but also can significantly improve the accuracy and robustness of the recommendation results through deep modeling of user positive and negative feedback information.

Claims

1. A social recommendation denoising method based on ternary resonance and polar modulation graph neural network is proposed herein, characterized in that, At least comprising the following steps: Step one: Perform priori preference consistency auditing of social links, by analyzing the ternary closed-loop structure formed by users' interactions with common items, quantifying preference consistency following the ternary relationship resonance principle, and generating a topological mask; Step two: Implement graph structure enhancement, using the topological mask to perform topological pruning on the original social network to achieve purification, and merging the purified high-confidence social relationships as structured priors with the user-item interaction graph; Step three: Examine the merged heterogeneous polarized network, and calculate the adaptive polarization modulation coefficients for each node; Step four: Perform bipolarity-aware graph convolutional representation learning, dynamically weighting the information flow in each polarized neighborhood in the graph neural network to learn node representations; Step five: Build a joint optimization framework and perform recommendations, optimizing the model by minimizing a composite objective function that integrates Bayesian personalized ranking loss and polarization order constraints, and ranking and recalling based on predicted preference scores.

2. The method of claim 1, wherein: The step one, the confidence assessment of each social relationship, first converts the user's rating data into a polarized bipartite graph representation to distinguish between positive and negative polarities of user preferences; second, extracts all shared interaction history by calculating two-hop paths between user nodes, and applies the ternary resonance principle to classify these paths into preference alignment or preference conflict; Finally, aggregate these classification results to build a global relationship confidence matrix that provides a priori judgment of each user pair based on their historical interaction behavior consistency, independent of any learning parameters.

3. The method of claim 1, wherein: The step two, the graph structure enhancement, first uses the relationship confidence matrix produced by the previous step as a logical filter to selectively retain the original connections of the social network, retaining only those social connections that are judged to be high-confidence, thereby filtering out inconsistent noise edges; Then, merge this purified social subgraph as a topological enhancement information layer with the adjacency matrix representation of the user-item bipartite graph through linear superposition, thereby outputting a unified multi-relational graph that integrates social homogeneity and user behavior patterns, providing a more information-rich and structurally reliable foundation for subsequent feature learning.

4. The method of claim 1, wherein: The step three, the calculation of modulation coefficients, first examines the structure of the heterogeneous information network by traversing each node in the network and analyzing the connection relationships within its local first-degree neighborhood to extract the number of positive and negative polarity neighbors as statistical features of the neighborhood environment; thereafter, dynamically calculates a unique pair of polarization modulation coefficients for each node based on the obtained neighborhood polarity statistical features, which are used to quantify the dependence on positive and negative polarity information streams in subsequent information aggregation.

5. The method of claim 1, wherein: The step four is used to learn the node representation. In an iterative neighborhood information propagation framework, the node-specific polarization modulation coefficient calculated in the previous step is used to differentially weight the incoming information from different polarity neighbors to achieve biased neighborhood information aggregation. Finally, the final entity latent feature vector that can capture multi-scale neighborhood structure is synthesized by combining the embedding generated in each iteration process using mean pooling.

6. The method of claim 1, wherein: In the step five, the optimization and recommendation are performed. First, a score estimation function is defined to estimate the affinity score by calculating the inner product of the user and item latent feature vectors. Second, a multi-component training criterion is developed, which integrates a pairwise ranking objective aimed at distinguishing observed positive interactions from unobserved samples, and a polarized ordinal constraint term for imposing explicit like-dislike boundaries, supplemented by a regularization term to control model complexity. Finally, the model weight set that makes the training criterion optimal is found by optimization algorithm, and a candidate list of items ranked in descending order of affinity is generated for the user by using the trained model for recommendation.

Images