A social recommendation method based on flow matching

Abstract

This invention discloses a social recommendation method based on flow matching. The method first extracts multi-scale basic features of users; then, it weights the social relationship matrix according to the overlap of interactive items between users, and constructs a social-aware preference manifold based on interaction features, using it as the target distribution for flow matching denoising; furthermore, it introduces a flow matching algorithm to progressively map noisy social features along the transmission path to the preference manifold space, and performs discrete updates during the transmission process to obtain denoised social features, achieving cross-domain feature alignment; finally, it dynamically fuses the denoised social features with interaction features to generate the final user representation, and calculates the recommendation score based on this representation. This invention explicitly corrects noise perturbations in social signals, effectively bridging the semantic gap between social relationships and user preferences, and significantly improves the accuracy and robustness of recommendation systems in data-sparse and high-noise environments.

Description

Technical Field

[0001] This invention belongs to the field of recommendation system technology, specifically relating to a social recommendation method based on flow matching. Background Technology

[0002] Recommender systems play a crucial role in alleviating information overload on modern digital platforms and improving user experience. However, traditional collaborative filtering methods generally face the severe challenge of data sparsity, meaning that most users have only a very small number of historical interaction records, making it difficult for the model to infer their reliable personalized preferences. To alleviate this problem, social recommendation has been proposed as an effective solution. Its core theoretical basis lies in social homogeneity and social influence, that is, users with social connections tend to have similar preferences, and users' interests are spread through social interactions.

[0003] In recent years, with the development of deep learning technology, social recommendation methods based on graph neural networks have gradually become the mainstream paradigm. These methods construct complex heterogeneous graph structures of user-item interactions and user-user social relationships, utilizing multi-layered message passing and iterative information aggregation mechanisms to enable personal feature representations to capture high-order neighborhood structures and complex social influences, thereby significantly improving the prediction accuracy of recommendation systems.

[0004] However, in real-world social network scenarios, directly utilizing observed social graphs for recommendation modeling still faces the following pressing technical challenges:

[0005] First, there are serious semantic gaps and social noise issues. Real-world social connections often stem from offline relationships with family and friends, colleagues, or social status. These connections do not necessarily translate into shared interest in specific items (such as goods or movies). Empirical studies show that there is a significant inconsistency between numerous social connections and users' actual preferences. Directly aggregating these unreliable social signals introduces a large amount of redundant or misleading noise into user representations, limiting the robustness of recommendation models.

[0006] Secondly, existing denoising methods have limitations. Traditional denoising methods often employ heuristic filtering or similarity-based hard truncation. These methods rely on discrete neighborhood aggregation, which easily disrupts the connectivity of the network topology and struggles to adapt to complex non-Gaussian noise distributions. Although recent studies have attempted to introduce generative models such as diffusion probability models or flow matching, these methods limit their denoising objectives to the social domain, i.e., they try to reconstruct clean social representations from noisy inputs. Because this type of "reductive denoising" process lacks explicit guidance from the interaction preference domain, the generated social features suffer from severe semantic misalignment between the geometric manifold and the real preference space.

[0007] Finally, the flow generation network itself suffers from spectral bias. In standard flow matching frameworks, multilayer perceptrons are typically used to fit continuous vector fields. According to spectral bias theory, the standard tends to capture smooth, low-frequency signals, making it difficult to fit cross-domain manifold mappings with drastic semantic changes. This results in the model generating overly smooth vector fields when faced with significant semantic differences between the social and interest domains, failing to achieve accurate feature rectification and alignment.

[0008] Therefore, how to break through the limitations of traditional "restorative noise reduction" and construct a high-precision cross-domain rectification mechanism that can explicitly bridge the semantic gap between social interaction and preferences and effectively overcome spectral bias is a core technical challenge that urgently needs to be solved in the current field of social recommendation noise reduction. Summary of the Invention

[0009] The purpose of this invention is to provide a social recommendation method based on flow matching, aiming to solve the problems of large social relationship noise interference, inconsistency between social information and user interest semantic space, and spectral bias in generative denoising models in existing social recommendation technologies. This invention constructs a socially-aware preference manifold as the target distribution for flow matching, introduces a corrected flow framework with a high-frequency signal amplification mechanism, and explicitly maps social features to the preference space, thereby achieving cross-domain manifold alignment and significantly improving the prediction accuracy and robustness of the recommendation system in data-sparse and high-noise environments.

[0010] A social recommendation noise reduction method based on stream matching and preference alignment, comprising the following steps:

[0011] Step 1, Dual-view network construction and multi-scale manifold encoding: Construct user social graphs using social relationships between users. Construct a user-item interaction graph using user and item interaction records. The first branch of the graph convolutional network is used to perform feature propagation on the user's social graph to extract an initial noisy social representation. The second branch of the graph convolutional network is used to perform feature propagation on the user-item interaction graph to extract interactive user representations. Representation of items .

[0012] Step 2: Construct the socially perceived preference manifold target distribution: Based on the overlap of common interactive items among users in the user-item interaction graph, reweight the social network matrix and combine it with the interactive user representations extracted in Step 1. Constructing a socially perceived preference manifold This is used as the target distribution for flow matching noise reduction.

[0013] Step 3: Cross-domain manifold alignment and denoising rectification based on flow matching: A flow matching algorithm is introduced. During the model training phase, a neural network is used to learn the connection between the initial noisy social representation and the continuous vector field of the target distribution. During the denoising inference phase, the time variable is amplified and encoded with high-frequency signals, and a discrete inverse update is performed using an ordinary differential equation solver to obtain the rectified and purified social features. .

[0014] Step 4: Dual-view feature fusion based on adaptive gating mechanism: Dynamically calculate the purified social features obtained in Step 3 using an adaptive gating mechanism. Interacting with users The fusion weights are used to generate the final comprehensive user representation. .

[0015] Step 5: Recommendation rating calculation and multi-task joint training optimization: Based on the comprehensive user representation generated in Step 4 The items extracted in step 1 represent The predicted preference score is calculated, and the model parameters are jointly optimized end-to-end by combining the Bayesian personalized ranking loss and the flow matching mean squared error loss. The recommendation result is then output based on the final predicted preference score.

[0016] Furthermore, in step 1, the second branch of the graph convolutional network is used to perform feature propagation on the user-item interaction graph to extract interactive user representations. Representation of items Specifically, in the user-item interaction graph In this process, the second branch of the graph convolutional network is used for feature propagation, and its first branch... The propagation rules of the layers are as follows: , in, The adjacency matrix of the interaction graph. This is the degree matrix corresponding to the interaction graph; This is the embedded representation of the (l+1)th layer of the interaction graph. The embedded representation of the l-th layer of the interaction graph; propagation to the l-th layer. After the layers are defined, the interactive user representation is obtained by summing and aggregating the embeddings of each layer. Representation of items ,in The set number of layers in the graph convolutional network; where and The first The user and item embedding representation after layer aggregation;

[0017] Similarly, the first branch of the graph convolutional network is used in the user social graph. Feature propagation is performed on the upper layer, the embeddings of each layer are extracted and summed to obtain the initial noisy social representation. .

[0018] Furthermore, the specific process of constructing the socially perceived preference manifold target distribution in step 2 includes:

[0019] Step 2.1: For the initial social network matrix Reweighting is performed; if two users with social links share interactive items, the initial social network matrix is ​​reweighted. The matrix entries are assigned the Jaccard similarity between the sets of items interacted by the two users; if there are no shared items, the value is set to a preset constant. This yields a reweighted social matrix;

[0020] Step 2.2: Based on the reweighted matrix, select the top-k similar neighbors of each user, preserve the edge relationships, and construct a reconstructed social matrix. ;

[0021] Step 2.3: Reconstruct the social matrix With the interactive user representation Multiplication generates the target distribution for flow matching denoising. .

[0022] Furthermore, the specific process of cross-domain manifold alignment and noise reduction rectification based on flow matching in step 3 includes:

[0023] Step 3.1: During the model training phase, construct the connection target distribution. Optimal transmission interpolation path with reference noise distribution , in For time variables, It is Gaussian distributed noise. For time step Path status at that time;

[0024] Step 3.2: Representation to interactive users Introducing noise disturbance to obtain ,in For standard normally distributed noise, The scaling factor is used to control the intensity of the disturbance. This indicates element-wise multiplication;

[0025] Step 3.3: In the training phase, extract the standard sinusoidal time-aware embedding. Optimal transmission interpolation path Time-aware embedding and the interactive features after perturbation The components are concatenated, and the dynamic weights of each component are calculated using an attention mechanism and then fused to obtain a joint representation. The input is fed into a neural network composed of multilayer perceptrons. In this process, the goal is to minimize the difference between the predicted target manifold and the true distribution. The mean squared error is used as the target for regression training;

[0026] Step 3.4: In the model denoising inference stage, in order to capture high-frequency signals, the time variable is... Multiply by the preset magnification factor Then, sinusoidal time encoding is performed to obtain the amplified time-aware embedding. ;

[0027] Step 3.5: In the denoising inference stage, set the initial state to a noisy social representation. Optimal transmission interpolation path The amplified time-aware embedding and interactive features Input to a trained neural network Predicting the target manifold And calculate the corresponding rectified velocity field. Using an Euler solver with a fixed step size Perform the following discrete inverse time update, where Total number of sampling steps: ; This represents the social features derived by backwards from the current time t, after a time step Δt, and the integral is used to inversely calculate the social features up to the current time t. At that time, the rectified embedding is obtained That is, the purified social characteristics .

[0028] Furthermore, the specific process of dual-view feature fusion based on adaptive gating mechanism in step 4 includes:

[0029] The interactive user representation With the purified social characteristics The system concatenates data and dynamically calculates user-level scalar gating weights using an adaptive gating mechanism. : ; in, This indicates a splicing operation. It is the Sigmoid activation function. Both are learnable parameter matrices and bias parameters;

[0030] The comprehensive user representation is obtained by weighted fusion using the following formula: .

[0031] Recommendation rating calculation and multi-task joint training optimization: Based on the comprehensive user representation generated in step 4 The items extracted in step 1 represent The predicted preference score is calculated, and the model parameters are jointly optimized end-to-end by combining the Bayesian personalized ranking loss and the flow matching mean squared error loss. The recommendation result is then output based on the final predicted preference score.

[0032] Step 5.1 Predicting Preference Rating Calculation: Obtaining the User Representation Matrix Corresponding user in China User feature vector and item representation matrix Corresponding items Item feature vector ; Calculate user data via vector inner product For items Predicted preference score Its formula is: Among them, superscript Represents the matrix transpose symbol;

[0033] Step 5.2: Calculate the recommendation ranking loss using Bayesian personalized ranking loss. The formula is ; in To train the set of triples, Indicates user A collection of historical items that have already been interacted with. For positive sample items, Negative sample items;

[0034] Step 5.3: Based on the preference target manifold predicted by the neural network during the training phase. The target distribution for noise reduction is matched with the constructed stream. Calculate the mean square error loss for stream matching: ; in, Represents the time variable and reference noise distribution Seeking expectations;

[0035] Step 5.4: Introduction Regularization terms are used to construct the final joint loss function for multiple tasks. For all trainable parameters in the model Perform end-to-end optimization: , in To match the loss weights to the multi-task balancing parameters of the control flow, is the regularization coefficient.

[0036] 3. Beneficial effects

[0037] Compared with the prior art, the present invention adopts the above technical solution and has the following advantages:

[0038] This invention achieves cross-domain manifold alignment, bridging the semantic gap between social and preference domains: existing denoising methods mainly operate within the social domain, making it difficult to fully align with the user preference space. This invention designs a cross-domain manifold alignment module, introducing a flow matching framework to learn a continuous transport mapping, directly rectifying the initial noisy social representation into a socially aware preference manifold, significantly enhancing the effectiveness and discriminative power of social embeddings in downstream recommendation tasks.

[0039] Overcoming the spectral bias of neural networks and effectively capturing high-frequency preference signals: The core innovation of this invention lies in strictly distinguishing the processing mechanism of time variables during model training and denoising inference stages. During inference denoising, a preset amplification factor is introduced to amplify and encode high-frequency signals of the time variables, overcoming the defect of ordinary multilayer perceptrons that bias towards low-frequency signals, leading to excessive smoothing of the vector field. This enables the model to accurately capture high-frequency, fine-grained preference features to cope with complex non-Gaussian noise.

[0040] Dynamic adaptive feature fusion enhances the model's robustness in sparse and noisy scenarios: This invention abandons the traditional simple average pooling and designs an adaptive gating mechanism. This mechanism can dynamically calculate the fusion weights of social and interaction features for different users, assigning high weights only when social priors provide valid information, thereby significantly reducing the interference of residual social noise. This method demonstrates excellent robustness in handling sparse data from cold-start users and complex structural noise from highly active users, significantly improving recommendation recall and ranking quality. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention.

[0042] Figure 2 This is a model diagram of a social recommendation method based on flow matching according to the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be described in detail below.

[0044] The following is only one embodiment of the present invention. The present invention has many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention. All such corresponding changes and modifications should fall within the protection scope of the appended claims.

[0045] This invention discloses a social recommendation method based on flow matching, which includes the following steps:

[0046] Step 1: Construct a user social graph using the social relationships between users. Construct a user-item interaction graph using user and item interaction records. The first branch of the graph convolutional network is used to perform feature propagation on the user's social graph to extract an initial noisy social representation. The second branch of the graph convolutional network is used to perform feature propagation on the user-item interaction graph to extract interactive user representations. Representation of items Traditional collaborative filtering methods suffer from generalization bottlenecks due to data sparsity constraints and single-view interaction representations. To introduce heterogeneous auxiliary information, a dual-view heterogeneous network architecture is constructed. Feature propagation is performed in parallel based on social network interaction networks, and heterogeneous view information is mapped to a decoupled feature space through multi-scale manifold encoding, providing a foundational multi-dimensional decoupled representation for subsequent cross-domain manifold alignment and conditional denoising.

[0047] Step 2: Construct the socially perceived preference manifold target distribution: Based on the overlap of commonly interacting items among users in the user-item interaction graph, reweight the edge weights of the social network matrix and combine them with the interactive user representations extracted in Step 1. Construct the target distribution for flow matching denoising. A significant semantic gap and unreliable noise perturbations exist between observed social connections and users' actual item preferences in real-world scenarios. This paper uses the overlap of shared interactions within the behavioral domain as a metric to conditionally constrain and quantitatively reweight a discrete initial social network matrix, achieving inverse calibration of social relationships based on behavioral preferences. By mapping the calibrated topology to the interactive user representations containing pure preference signals, a socially perceived preference manifold target distribution is explicitly constructed. This distribution not only possesses the prior connectivity of the social graph but also aligns geometrically with the collaborative preference space, thus providing a cross-domain guided target distribution for subsequent flow matching denoising tasks. This fundamentally avoids the semantic misalignment in the geometric manifold inherent in traditional reductive denoising.

[0048] Step 3: Cross-domain manifold alignment and denoising rectification based on flow matching: A flow matching algorithm is introduced. During the model training phase, a neural network is used to learn the connection between the initial noisy social representation and the continuous vector field of the target distribution. During the denoising inference phase, the time variable is amplified and encoded with high-frequency signals, and a discrete inverse update is performed using an ordinary differential equation solver to obtain the rectified and purified social features. The original social connections, stemming from offline social relationships, suffer from a significant semantic gap and topological noise compared to real-world item preferences. Furthermore, traditional generative denoising methods are largely confined to reconstruction within the social domain, lacking explicit cross-domain guidance, and standard neural networks exhibit spectral bias towards low-frequency signals when faced with drastic cross-domain semantic changes. To address this, this step introduces a continuous flow matching framework. During the training phase, a deterministic ordinary differential equation mapping is established by learning to connect the noisy initial social representation to a continuous vector field representing the target distribution of the socially perceived preference manifold. During the inference phase, a pre-defined amplification factor is introduced to amplify and encode the time variable at high frequencies, enhancing the network's ability to capture fine-grained, high-frequency preference features. Finally, an Euler solver is used to perform discrete inverse trajectory integration along the rectified velocity field, accurately advancing the initial noisy social representation into the preference space.

[0049] Step 4: Dual-view feature fusion based on adaptive gating mechanism: Dynamically calculate the purified social features obtained in Step 3 using an adaptive gating mechanism. Interacting with users The fusion weights are used to generate the final comprehensive user representation. This mechanism can nonlinearly and adaptively adjust the sparsity of behavior for different individuals in the global parameter space, giving higher fusion weights to effective priors, thereby significantly suppressing possible residual noise interference and improving the generalization robustness of the overall model in multiple scenarios.

[0050] Step 5: Recommendation rating calculation and multi-task joint training optimization: Based on the comprehensive user representation generated in Step 4 The items extracted in step 1 represent The system calculates predicted preference scores, and performs end-to-end multi-task joint optimization of model parameters by combining Bayesian personalized ranking loss and flow matching mean squared error loss. Recommendation results are output based on the final predicted preference scores. The ultimate goal of the recommendation system is to optimize the evaluation metric, while the core objective of flow matching is to minimize the fitting error of the continuous vector field. To ensure that the purified manifold features can accurately adapt to downstream personalized preference predictions, a joint loss optimization function is constructed for multi-objective optimization.

[0051] In this embodiment, step 1 specifically includes the following steps:

[0052] Feature propagation is performed on the user-item interaction graph to extract interactive user representations. Representation of items Specifically, feature propagation is performed using the second branch of the graph convolutional network, where the first branch... The propagation rules of the layers are as follows: , in, The adjacency matrix of the interaction graph. This is the degree matrix corresponding to the interaction graph. This is the embedded representation of the (l+1)th layer of the interaction graph. This is the embedded representation of the l-th layer of the interaction graph.

[0053] Spread to the After the layers are defined, the interactive user representation is obtained by summing and aggregating the embeddings of each layer. Representation of items ,in The set number of layers in the graph convolutional network; where and The first The user and item embeddings are then aggregated into layered representations; similarly, in the user social graph... The first branch is used to perform independent message passing operations, extract the embeddings of each layer and sum them to obtain the initial noisy social representation. .

[0054] In this embodiment, step 2 specifically includes the following steps:

[0055] Step 2.1: For the initial social network matrix Reweighting is performed; if two users with social links share interactive items, the initial social network matrix is ​​reweighted. The matrix entries are assigned the Jaccard similarity between the sets of items interacted by the two users; if there are no shared items, the value is set to a preset constant. This yields the reweighted social matrix. Preset constants are used. As a smoothing term and lower bound for reweighting, if two users share a social connection but their behavior is extremely sparse, resulting in a calculated item overlap of 0, directly truncating it to zero would completely lose this valuable social prior; however, by assigning it a small positive number... This mechanism not only distances itself from systems that are "lacking both social interaction and engagement" but also maintains basic connectivity during streaming matching. It effectively alleviates the cold start and data sparsity problems faced by recommendation systems.

[0056] Step 2.2: Based on the reweighted matrix, select each user's... Preserve edge relationships among similar neighbors to construct and reconstruct the social matrix. ;

[0057] Step 2.3: Reconstruct the social matrix With the interactive user representation Multiplication generates the target distribution for flow matching denoising. .

[0058] In this embodiment, step 3 specifically includes the following steps:

[0059] Step 3.1: During the model training phase, construct the connection target distribution. Optimal transmission interpolation path with reference noise distribution , in For time variables, It is Gaussian distributed noise. For time step Path status at that time;

[0060] Step 3.2: Representation to interactive users Introducing noise disturbance to obtain ,in For standard normally distributed noise, The scaling factor is used to control the intensity of the disturbance. This indicates element-wise multiplication;

[0061] Step 3.3: In the training phase, extract the standard sinusoidal time-aware embedding. Optimal transmission interpolation path Time-aware embedding and the interactive features after perturbation The components are concatenated, and the dynamic weights of each component are calculated using an attention mechanism and then fused to obtain a joint representation. The input is fed into a neural network composed of multilayer perceptrons. In this process, the goal is to minimize the difference between the predicted target manifold and the true distribution. The mean squared error is used as the target for regression training;

[0062] Step 3.4: In the model denoising inference stage, in order to capture high-frequency signals, the time variable is... Multiply by the preset magnification factor Then, sinusoidal time encoding is performed to obtain the amplified time-aware embedding. ;

[0063] Step 3.5: In the denoising inference stage, set the initial state to a noisy social representation. Optimal transmission interpolation path The amplified time-aware embedding and interactive features Input to a trained neural network Predicting the target manifold And calculate the corresponding rectified velocity field. Using an Euler solver with a fixed step size Perform the following discrete inverse time update, where Total number of sampling steps: ; This represents the social features derived by backwards from the current time t, after a time step Δt, and the integral is used to inversely calculate the social features up to the current time t. At that time, the rectified embedding is obtained That is, the purified social characteristics .

[0064] In this embodiment, step 4 specifically includes the following steps: representing the interactive user... With the purified social characteristics The system concatenates data and dynamically calculates user-level scalar gating weights using an adaptive gating mechanism. : , in, This indicates a splicing operation. for Activation function All parameters are learnable parameter matrices and bias parameters; the comprehensive user representation is obtained by weighted fusion using the following formula: .

[0065] In this embodiment, step 5 specifically includes the following steps:

[0066] Recommendation rating calculation and multi-task joint training optimization: Based on the comprehensive user representation generated in step 4 The items extracted in step 1 represent Calculate the predicted preference score, perform end-to-end multi-task joint optimization of the model parameters by combining the Bayesian personalized ranking loss and the streaming matching mean squared error loss, and output the recommendation result based on the final predicted preference score. Step 5.1: Obtain the user representation matrix Corresponding user in China User feature vector and item representation matrix Corresponding items Item feature vector ; Calculate user data via vector inner product For items Predicted preference score Its formula is: Among them, superscript Represents the matrix transpose symbol;

[0067] Step 5.2: Calculate the recommendation ranking loss using the Bayesian personalized ranking loss formula: , in To train the set of triples, Indicates user A collection of historical items that have already been interacted with. For positive sample items, Negative sample items;

[0068] Step 5.3: Based on the preference target manifold predicted by the neural network during the training phase in Step 3.3. The target distribution for noise reduction is matched with the constructed stream. Calculate the mean square error loss for stream matching: ; in, Represents the time variable and reference noise distribution Seeking expectations;

[0069] Step 5.4: Introduction Regularization terms are used to construct the final joint loss function for multiple tasks. , in, To match the parameters of the loss weights to the control flow, The regularization coefficient is used. The m-optimization algorithm trains the model to obtain the parameters that best perform the model.

[0070] Simulation Experiment: The experimental environment for this invention is equipped with... 64-bit operating system machine, for ,

[0071] The memory capacity was 32GB. Finally, experiments were conducted on three public datasets (Ciao, Epinions, and Yelp). The dataset statistics are as follows:

[0072] Table 1 Data Statistics

[0073] To accurately evaluate the recommendation prediction accuracy and ranking quality of the method described in this invention, the simulation experiments employed two internationally recognized recommendation system evaluation metrics and conducted in-depth comparative tests with more than ten mainstream baseline models. For clarity, the English abbreviations and their Chinese meanings used in the experimental results table are explained below. Among the evaluation metrics, (Recall rate@20) represents the percentage of items that users actually interacted with among the top 20 candidate recommended items output by the system, used to evaluate the completeness of the model's recall of target preferred items; (Normalized loss cumulative gain @20) is a multi-level ranking accuracy metric that considers positional weights and is used to evaluate the quality of the item preference ranking output after noise reduction and rectification. In the baseline model, For a matrix factorization model based on trust relationships, trust topology is injected into collaborative filtering; To create a social influence diffusion network, interest diffusion is simulated hierarchically on the original social graph; This is a graph neural network-based social recommendation model that aggregates social views and interaction signals. For dynamic graph attention-based social recommendation models, capture the dynamic influence of neighborhood in the behavioral conversation flow; This is a neural graph collaborative filtering model that mines higher-order connectivity by passing collaborative signals on the interaction graph. For a multi-channel hypergraph convolutional social recommendation network, social homogeneity is captured by modeling high-order primitive structures such as social triangles. A knowledge-aware coupled graph neural network model is used to link external knowledge and multi-dimensional user social interactions for message passing. This is a self-supervised metagraph information maximization network that utilizes metagraph structure to extract heterogeneous features and maximizes mutual information through self-supervised contrast constraints. To guide a graph-guided differentiable social recommendation model, a differentiable relation pruning strategy is used to attempt to filter local edge noise. To denoise the self-enhancing social recommendation learning model, inconsistent information is filtered out by constructing a two-end self-supervised contrast constraint. To provide a social recommendation noise reduction framework based on a diffusion probability model, a diffusion model is introduced to perform restorative representation purification within the social domain. This is a social recommendation denoising framework based on flow matching, which utilizes standard flow matching to perform denoising within the social domain.

[0074] The core of this invention is the method (MARF, a social recommendation denoising model combining flow matching and manifold alignment). It constructs a socially aware preference manifold as the cross-domain target distribution and uses a corrected flow framework to denoise cross-manifold features. Combining the aforementioned evaluation metrics and the rigorous technical definitions of the model method, cross-domain alignment and robust prediction tests were conducted on three real-world public datasets. The specific recommendation performance results are shown in Table 2. Experimental data show that the method performs excellently on datasets of different sizes, especially on the Epinions dataset, which has significant social attributes and noise. Compared to the strongest baseline model, RecFlow, this method achieves a significant improvement of 5.56% in recall. This fully verifies that by constructing a socially aware preference manifold and combining it with a corrected flow framework with a high-frequency signal amplification mechanism, the semantic gap between the social domain and the preference domain can be explicitly bridged, effectively removing redundant social noise while significantly improving the accuracy and robustness of the recommendation results.

[0075] Table 2 shows the comparison results of the method of this invention with other methods on three datasets:

Claims

1. A social recommendation noise reduction method based on stream matching, characterized in that, The method includes the following steps: Step 1, Dual-view network construction and multi-scale manifold encoding: Construct user social graphs using social relationships between users. Construct a user-item interaction graph using user and item interaction records. The first branch of the graph convolutional network is used to perform feature propagation on the user's social graph to extract an initial noisy social representation. The second branch of the graph convolutional network is used to perform feature propagation on the user-item interaction graph to extract interactive user representations. Representation of items ; Step 2: Construct the socially perceived preference manifold target distribution: Based on the overlap of common interactive items among users in the user-item interaction graph, reweight the social network matrix and combine it with the interactive user representations extracted in Step 1. Constructing a socially perceived preference manifold , and use it as the target distribution for flow matching noise reduction; Step 3: Cross-domain manifold alignment and denoising rectification based on flow matching: A flow matching algorithm is introduced. During the model training phase, a neural network is used to learn the connection between the initial noisy social representation and the continuous vector field of the target distribution. During the denoising inference phase, the time variable is amplified and encoded with high-frequency signals, and a discrete inverse update is performed using an ordinary differential equation solver to obtain the rectified and purified social features. ; Step 4: Dual-view feature fusion based on adaptive gating mechanism: Dynamically calculate the purified social features obtained in Step 3 using an adaptive gating mechanism. Interacting with users The fusion weights are used to generate the final comprehensive user representation. ; Step 5: Recommendation rating calculation and multi-task joint training optimization: Based on the comprehensive user representation generated in Step 4 The items extracted in step 1 represent The predicted preference score is calculated, and the model parameters are jointly optimized end-to-end by combining the Bayesian personalized ranking loss and the flow matching mean squared error loss. The recommendation result is then output based on the final predicted preference score.

2. The social recommendation noise reduction method based on stream matching according to claim 1, characterized in that, In step 1, the second branch of the graph convolutional network is used to perform feature propagation on the user-item interaction graph to extract interactive user representations. Representation of items Specifically, in the user-item interaction graph In this process, the second branch of the graph convolutional network is used for feature propagation, and its first branch... The propagation rules of the layers are as follows: in, The adjacency matrix of the interaction graph. This is the degree matrix corresponding to the interaction graph. This is the embedded representation of the (l+1)th layer of the interaction graph. This is the embedded representation of the l-th layer of the interaction graph; Spread to the After the layers are defined, the interactive user representation is obtained by summing and aggregating the embeddings of each layer. Representation of items ,in The set number of layers in the graph convolutional network; and The first The user and item embedding representation after layer aggregation; Similarly, the first branch of the graph convolutional network is used in the user social graph. Feature propagation is performed on the upper layer, the embeddings of each layer are extracted and summed to obtain the initial noisy social representation. .

3. The social recommendation noise reduction method based on stream matching according to claim 2, characterized in that, The specific process of constructing the socially perceived preference manifold target distribution in step 2 includes: Step 2.1: For the initial social network matrix... Reweighting yields a reweighted social matrix. During the reweighting process, if two users with social links share interactive items, the initial social network matrix will be reweighted. The matrix entries are assigned the Jaccard similarity between the sets of items interacted by the two users; if there are no shared items, the value is set to a preset constant. ; Step 2.2: Based on the reweighted social matrix obtained in Step 2.1, select the top-k similar neighbors of each user, retain the edge relationships, and construct a reconstructed social matrix. ; Step 2.3: Reconstruct the social matrix With the interactive user representation Multiplication generates a socially perceived preference manifold. , which serves as the target distribution for flow matching noise reduction.

4. The social recommendation noise reduction method based on stream matching according to claim 3, characterized in that, The specific process of cross-domain manifold alignment and noise reduction rectification based on flow matching in step 3 includes: Step 3.1: Construct the forward propagation path: During the model training phase, construct the connection target distribution. Optimal transmission interpolation path with reference noise distribution ,in For time variables, It is Gaussian distributed noise. For time step Path status at that time; Step 3.2: Introduce feature uncertainty perturbation: for interactive user representation Introducing noise perturbation to obtain the perturbed interaction characteristics ,in For standard normally distributed noise, The scaling factor is used to control the intensity of the disturbance. This indicates element-wise multiplication; Step 3.3, Vector Field Learning in the Training Phase: In the training phase, extract the standard sinusoidal time-aware embedding. Optimal transmission interpolation path Standard sinusoidal time-aware embedding and the interactive features after perturbation The components are concatenated, and the dynamic weights of each component are calculated using an attention mechanism and then fused to obtain a joint representation. The input is fed into a neural network composed of multilayer perceptrons. In this process, the goal is to minimize the difference between the predicted target manifold and the true distribution. The mean squared error is used as the target for regression training; Step 3.4, Time Signal Amplification in the Inference Stage: In the model denoising inference stage, the time variable is amplified... Multiply by the preset magnification factor Then, sinusoidal time encoding is performed to obtain the amplified time-aware embedding. ; Step 3.5, Discrete Inverse Rectification Update: In the denoising inference stage, the initial state is set to the noisy social representation. Optimal transmission interpolation path The amplified time-aware embedding and the interactive features after perturbation Input to a trained neural network Predicting the target manifold And calculate the corresponding rectified velocity field. Using an Euler solver with a fixed step size Perform the following discrete reverse time update, where Total number of sampling steps: This represents the social features derived by backwards from the current time t, after a time step Δt, and the integral backwards to... At that time, the rectified embedding is obtained That is, the purified social characteristics .

5. The social recommendation noise reduction method based on stream matching and preference alignment according to claim 4, characterized in that, The specific process of dual-view feature fusion based on adaptive gating mechanism in step 4 includes: The interactive user representation With the purified social characteristics The system concatenates data and dynamically calculates user-level scalar gating weights using an adaptive gating mechanism. : in, This indicates a splicing operation. It is the Sigmoid activation function. Both are learnable parameter matrices and bias parameters; A comprehensive user representation is obtained through weighted fusion: .

6. The social recommendation noise reduction method based on stream matching and preference alignment according to claim 5, characterized in that, The specific process of recommending score calculation and multi-task joint training optimization in step 5 includes: Step 5.1 Predicting Preference Rating Calculation: Obtaining the User Representation Matrix Corresponding user User feature vector and item representation matrix Corresponding items Item feature vector ; Calculate user data via vector inner product For items Predicted preference score Its formula is: , where superscript Represents the matrix transpose symbol; Step 5.2 Calculate the recommendation ranking loss The Bayesian personalized ranking loss is calculated using the following formula: in To train the set of triples, Indicates user A collection of historical items that have already been interacted with. For positive sample items, For negative sample items, For users For items Predicted preference scores; Step 5.3 Calculate the mean square error loss for stream matching. Based on the predicted target manifold obtained in step 3.5 The socially perceived preference manifold distribution constructed in step 2.3 Calculate the mean square error: in, Represents the time variable and reference noise distribution Seeking expectations; Step 5.4 Multi-task joint loss optimization: Introducing Regularization terms are used to construct the final joint loss function for multiple tasks. For all trainable parameters in the model Perform end-to-end optimization: in, To match the loss weights to the multi-task balancing parameters of the control flow, The regularization coefficient is used. Based on the loss function mentioned above, the Adam optimization algorithm is used to train the parameters that make the model perform optimally.

Images