Recommendation method for explicitly dissociating social contact and sequence factors
By explicitly dissociating social and sequential factors through the multi-channel graph propagation module and cross-view comparative learning module, the problem of difficulty in separating social and sequential influences in existing recommendation systems is solved, achieving more accurate and personalized recommendation effects.
Patent Information
- Application Number
- CN202510872294.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-10-03
AI Technical Summary
Existing recommendation systems find it difficult to explicitly dissociate social and sequential factors, fail to capture high-order influences, and have difficulty dynamically distinguishing individual differences among users, resulting in insufficient recommendation accuracy and interpretability.
A multi-channel graph propagation module, a cross-view comparative learning module, and a model prediction optimization module are used to capture high-order information of social and sequential influences respectively, and to perform deep dissociation and dynamically regulate the weights of influencing factors through self-supervisory signals.
The prediction accuracy and interpretability of the recommendation system are improved, especially the recommendation effect in inactive users and high-order relationship scenarios, and dynamically adapt to individual differences of users.
Smart Images
Figure CN120745825A_ABST
Abstract
Description
[0001] Technology Space
[0002] The present invention relates to the fields of machine learning and information retrieval, and in particular to a recommendation method that explicitly dissociates social and sequential factors. Background Art
[0003] Recommender systems play an important role in alleviating information overload. Their primary goal is to recommend content tailored to users' interests and preferences. Consequently, they are widely used in a variety of fields, including e-commerce, location-based services, and short video platforms. The core of personalized recommendations lies in leveraging historical interactions to accurately predict items of interest to users. Traditional recommendation systems are primarily based on collaborative filtering, and their development can be roughly divided into three stages: early methods based on nearest neighbors and matrix decomposition, mid-term neural collaborative filtering methods, and more recently, graph neural recommendation methods. Although these methods have made significant progress, collaborative filtering-based methods typically only capture static user preferences and struggle to capture dynamic user interests, thus limiting recommendation performance.
[0004] Compared to traditional collaborative filtering-based recommendation methods, sequential recommendation better captures the dynamic drift of user interests by modeling sequential patterns between items. Early research primarily employed decomposition-based personalized Markov chain methods, extracting users' short-term interests through first-order Markov chains. Subsequently, many studies extended this approach to higher-order Markov chains, finding that mining sequential patterns between multiple items recently visited by a user further enhanced the model's recommendation effectiveness. With the resurgence of deep learning, subsequent research has gradually shifted to designing sequential recommendation models using various deep neural networks, including those based on recurrent neural networks, convolutional neural networks, transformers, and graph neural network recommendation. Graph neural network recommendation methods have garnered widespread attention in recent years because graph structures are naturally well-suited to capturing complex item transfer information. However, these methods primarily focus on user behavior and ignore the social influence between different users, limiting their applicability to scenarios with inactive users.
[0005] Recent research has shown that modeling the influence of users' social relationships is as important as the influence of their behavioral sequences for predicting items of interest. For example, conversation-aware recommendation methods based on graph neural networks utilize heterogeneous graph neural networks to capture static social influences and learn session-specific representations through gated graph neural networks. While existing research has achieved some results, it still suffers from the following three shortcomings:
[0006] (1) Failure to explicitly disentangle social and sequential factors. Although the above methods take social influence into account when modeling user sequences, they usually model these two influences in a coupled manner. To explain this limitation, Figure 1A simple example from a point-of-interest recommendation scenario is presented: Analyzing solely from a sequential or social perspective makes it difficult to determine whether user u1's move from location i4 to location i5 was influenced by sequential patterns or driven by social connections. In reality, user u1's habit of visiting the stadium after dining aligns with both the item transfer pattern in the sequence and the historical trajectory of his friend u2 in his social network. The combined influence of these two factors makes it difficult to accurately interpret the user's true intent. Therefore, explicitly separating and modeling social and sequential influences is crucial for improving the predictive accuracy and interpretability of recommendation systems.
[0007] (2) Failure to capture the high-order influence of social relationships and sequence patterns. When predicting sparse user behavior, only modeling the low-order information of social and sequence influences may produce significant bias. Figure 1 , when user u4 wants to find a fitness place, if we fully consider the high-level information u1→u2→u3→u4 in the graph, we can recommend gym i1 to him. At the same time, if we recommend the next point of interest to user u2 with a limited visit sequence, it is difficult to make an accurate prediction. However, if we integrate the sequences of all users into an item sequence graph, we can find that there is Then, we can reasonably recommend cafe i7 to user u2 who has visited cafes i6, i4, and i5. Therefore, explicitly capturing the high-order influence of social relationships and sequential patterns helps better model user preferences.
[0008] (3) It is difficult to dynamically distinguish the differentiated effects of various influencing factors. After completing the decomposition of social and sequential influence, if only a fixed aggregation strategy (such as a simple splicing or summation operation) is used to integrate the influence representations at different levels, it will be difficult to reflect the differences between individual users. For example, extroverted users may be more inclined to adopt friends' suggestions and explore novel experiences such as farm stays; while users who like to be alone may choose fast food restaurants that are closer. Therefore, designing a dynamic and adaptive influencing factor aggregation mechanism to more finely measure the personalized differences of different users is of vital importance for accurate personalized recommendations. Summary of the Invention
[0009] In response to the above problems, the present invention discloses a recommendation method that explicitly dissociates social and sequential factors, which includes three core components: a multi-channel graph propagation module, a cross-view comparative learning module, and a model prediction and optimization module. First, the present invention decomposes the complex interaction graph into a user social graph and an item sequence graph, and designs a multi-channel graph propagation module to capture the high-order information of social and sequential influences in each graph structure. Secondly, a cross-view comparative learning module is constructed, and the social and sequential information learned in the user-item bipartite graph is used as a self-supervisory signal to achieve a deep dissociation of the two influencing factors. Finally, a personalized control mechanism is designed to dynamically adjust the weights of different influencing factors in model prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in examples or descriptions of the prior art.
[0011] Figure 1 This is an example diagram of a recommended scenario of the present invention;
[0012] Figure 2 It is a schematic diagram of the DSSRec model framework of the present invention;
[0013] Figure 3 This is a comparison chart of the experimental results of the present invention under different recommendation lists;
[0014] Figure 4 This is a comparison chart of the experimental results of the present invention under different user groups;
[0015] Figure 5 This is a comparison chart of the experimental results of the present invention under different numbers of graph propagation layers;
[0016] Figure 6 It is a comparison chart of the training time of the present invention and the benchmark method.
[0017] Specific implementation
[0018] 1. Problem Definition
[0019] Definition 1 User Social Graph
[0020] make Represents a collection of users, Represents the relationship matrix between users and defines Represents the friend set of each user. This constructs the social graph of all users in Represents the edge set of the node, directed edge (u, v)∈ε s Indicates that user u trusts user v.
[0021] Definition 2 Item Sequence Diagram
[0022] set up Represents a collection of items, Represents the set of interactive items of user u. According to the order of interaction, the interaction sequence sorted by time is obtained The interaction sequences of all users constitute a set Construct an item sequence diagram from this Among them, directed edges Indicates items To Item The transfer relationship, edge type γ u Determined by user information and time interval information.
[0023] Definition 3: User-Item Bipartite Graph
[0024] make represents the user-item interaction matrix, where r ui Indicates whether user u has interacted with item i. Based on this matrix, a user-item bipartite graph is constructed. where the edge set ε I Represents the interaction relationship between any two nodes in the graph.
[0025] Definition 3: Socially Aware Sequential Recommendation
[0026] Input: Given a user behavior sequence User social graph G S , Item Sequence Diagram G H , user-item bipartite graph G I , friends gathering And the set of items that the user has visited in the test set
[0027] Output: For each user u, generate an ordered list of item recommendations Specifically:
[0028]
[0029] where Prob(·) represents the probability that user u visits item i.
[0030] 2. Overall Model Architecture
[0031] Figure 2 The overall architecture of the DSSRec model of the present invention is presented. The model consists of three components: a multi-channel graph propagation module, a cross-view contrastive learning module, and a model prediction and optimization module. These modules are described in detail below:
[0032] 2.1 Multi-channel graph propagation module
[0033] Social relationships and behavioral sequences are two key factors that influence user decisions. In order to effectively distinguish and model the independent influences at the social and sequence levels, this paper designs a multi-channel graph propagation module consisting of a representation initialization layer and three graph convolutional layers.
[0034] (1) Characterization initialization layer
[0035] First, initialize two independent representation vectors for users and items respectively, where They serve as the social initialization representation and interest initialization representation of user u respectively; They represent the sequence initialization representation and interest initialization representation of item i respectively. The initial representation matrix of all users and items is as follows:
[0036]
[0037] (2) Social Perception Graph Communication Layer
[0038] Addressing the limitation of traditional graph neural network propagation functions, which generally assume that all neighbor nodes contribute equally to the target node, this paper proposes a propagation function that considers the heterogeneity of social influence, more accurately reflecting the reality that different friends have different influences on the target user. Given that friends with similar interests tend to have a more significant influence on users, this method can more accurately identify and model high-level structural information in the user's social graph. Its formal definition is as follows:
[0039]
[0040] Among them, α uv represents the social influence weight of target user u and friend v, and s(·,·) represents the cosine similarity function. S After layer propagation, the representation set of user u at each layer can be obtained Then, average pooling is used to aggregate the representations of each layer to obtain the final representation of the user, which is as follows:
[0041]
[0042] (3) Sequence Perception Graph Propagation Layer
[0043] Existing research on sequential recommendation is primarily limited to modeling single user behavior sequences and fails to fully exploit collaborative information across sequences. To this end, this paper designs a sequence-aware graph propagation layer to effectively model high-order transfer patterns in item sequence graphs. Specifically, a sequence graph is first constructed based on all user sequences, with items as nodes and users as edge type annotations. In this sequence graph, multiple types of directed edges may exist between any two nodes, representing item transfer relationships between different users. The information propagation process from item i to item j via user u can be formally expressed as follows:
[0044]
[0045] in represents the set of items preceding item j, and ε represents the edge set from item i to item j. The present invention constructs a time interval-aware transfer mechanism to calculate the key problem Specifically, calculate the time interval representation of each user's previous and subsequent access sequences This gives the transfer vector: final, The calculation form is as follows:
[0046]
[0047] Perform K on sequence diagram H After propagation, we can get the representation set of all layers of items Afterwards, average pooling is used to obtain the final representation of the item:
[0048]
[0049] (4) Interest-aware graph propagation layer
[0050] Based on the structural characteristics of the user-item bipartite graph, this paper uses lightweight graph convolution operations to capture collaborative information in the graph. The message propagation rule between user u and item i can be formally defined as follows:
[0051]
[0052] Perform K on the user-item bipartite graph I After the message propagation, the final representation of users and items in the user interest space is generated through average pooling. The specific calculation formula is as follows:
[0053]
[0054] 2.2 Cross-view contrastive learning module
[0055] To facilitate the deep decomposition of social and sequential influence, this paper uses the social and sequential information learned from the user-item bipartite graph as self-supervisory signals to optimize the social and sequential influence representations learned from the social graph and the social graph, respectively. Therefore, we first use the attention mechanism to calculate the social influence representation of user u and the friend preference representation in the user-item bipartite graph:
[0056]
[0057] in, is a learnable parameter. In addition, the present invention introduces contrastive learning to enhance the consistency between the user's social influence and the friend's preference:
[0058]
[0059] Among them, τ is a temperature hyperparameter that adjusts the contrastive learning scale. Similarly, the sequence influence representation and sequence proxy representation are calculated:
[0060]
[0061] in, is a learnable parameter. The corresponding contrastive learning objective is calculated as follows:
[0062]
[0063] 2.3 Model Prediction
[0064] After executing the multi-channel graph propagation module and the cross-view contrastive learning module, the user’s social influence representation can be obtained. and sequence-affected representation In order to adaptively fuse these two representations, the present invention designs an adaptive allocation mechanism based on self-attention, which enables the model to dynamically distinguish the contribution weights of different influencing factors to user representation. The calculation form is as follows:
[0065]
[0066] in, is the weight parameter. Then, the attention weight is normalized:
[0067]
[0068] Based on the normalized attention weight value, the user u’s attention to the item can be generated The preference score is calculated as follows:
[0069]
[0070] in Used to capture items Popular deviation.
[0071] 3. Model Optimization
[0072] The present invention adopts the sequence Bayesian personalized ranking loss function as the main optimization target of the model. Indicates that user u accessed the item at time t And through random sampling strategy, an item j that the user has not interacted with is selected as a negative sample. Based on the Bayesian personalized ranking hypothesis, user u is more inclined to choose item The conditional probability of non-item j can be formally expressed as:
[0073]
[0074] Where Θ represents all learnable parameters of the model. Then, by minimizing the negative log-likelihood function, the optimization objective can be obtained:
[0075]
[0076] Where λ is the regularization coefficient. Finally, the overall loss function of the model is a weighted combination of S-BPR loss and dual contrast learning loss:
[0077]
[0078] Table 1 details the model training process.
[0079] 3.1 Complexity Analysis
[0080] (1) Space complexity
[0081] As shown in Table 1, the spatial complexity of the model is mainly determined by the parameter set Θ. Among them, Θ1 represents two sets of initialized user representations and sequence representations, and its spatial complexity is It is linearly related to the number of users and the number of items. Θ2 is a learnable weight parameter with a space complexity of (8+6D)D. Since the hyperparameter D is much smaller than and Its storage cost is almost negligible.
[0082] (2) Time complexity
[0083] The computational overhead of the present invention mainly comes from the following components: social perception graph propagation layer, sequence perception graph propagation layer, interest perception graph propagation layer, and cross-view comparative learning module. Specifically, the time complexity of the three graph propagation layers is determined by the number of propagation layers K and the time complexity of the three graph propagation layers K. s , K H and K I Decision, in order and The time complexity of the cross-view contrastive learning module is in represents the average number of users. In summary, the total time complexity of the DSSRec model is
[0084] Figure 3 The accuracy comparison between the proposed method and the baseline method at different recommendation list lengths is shown; Figure 4 The comparison of normalized loss-cumulative gain of item ranking under different user groups of the present invention is demonstrated; Figure 5We present a hyperparameter sensitivity analysis of the invention with respect to the number of graph propagation layers. Figure 6 The efficiency and time performance of our proposed method were compared with those of a baseline method. We propose a recommendation method, DSSRec, that explicitly disentangles social and sequential factors. This method achieved excellent performance in the aforementioned experiments, outperforming existing work in terms of recommendation effectiveness and fully demonstrating the rationality and effectiveness of decomposing and modeling high-order social relationships and sequential influences. Future work will explore the integration of this method with efficient multimodal data to alleviate the ubiquitous problem of data sparsity, thereby further improving recommendation accuracy.
[0085] Those skilled in the art will readily understand the embodiments of the present invention. The above description is merely an example of a basic implementation of the present invention and is not intended to limit the present invention.
[0086] Table 1 Optimization process of DSSRec model
[0087]
Claims
1. A recommendation method that explicitly dissociates social and sequential factors, characterized in that: (1) Construct a representation initialization module to separately model the social relationships of users and the sequential behaviors of items, and generate user social graphs and item sequence graphs respectively; (2) Design a multi-channel graph propagation module, including a social-aware graph propagation layer and a sequence-aware graph propagation layer, to respectively characterize the social influence between users and the sequential patterns between items, thereby effectively capturing the high-order correlation information in the two graph structures; (3) Combine the user-item bipartite graph to extract social and sequential information as self-supervisory signals, and introduce a double contrast learning mechanism to achieve deep dissociation of the two influences; (4) Construct a personalized regulation mechanism to dynamically distinguish the importance of different influencing factors during prediction.
2. A recommendation method for explicitly dissociating social and sequential factors according to claim 1, characterized in that: The data input of the model is the user behavior sequence User social graph G S , Item Sequence Diagram G H , user-item bipartite graph G I and friends gathering 3. The recommendation method according to claims 1 and 2, characterized in that In the representation initialization module, in order to better achieve the decomposition learning of social influence and sequence influence, all users and items are initialized into two sets of independent representation matrices:
4. The recommendation method according to claims 1, 2 and 3, characterized in that In the social-aware graph propagation module, considering that different friends have different influences on the target user, and friends with similar interests often have a more significant influence on the user, a personalized social influence propagation function is used to effectively capture high-level information in the user's social graph: And, after K S After layer propagation, average pooling is used to aggregate the representations of each layer to obtain the final representation of the user:
5. The recommendation method according to claims 1, 2 and 3, characterized in that In the sequence-aware graph propagation module, to fully capture cross-sequence collaborative information, a time interval-aware graph information propagation mechanism is constructed to effectively model the high-order transfer patterns between items in all user sequences. Specifically: Calculate the time interval representation of each user's previous and subsequent access sequences Get the transfer vector: Then, the information propagation process of item i being transferred to item j through user u can be expressed as: By K H After propagation, we get the representation set of all layers Finally, average pooling is used to obtain the final representation of the item:
6. The recommendation method according to claims 1, 2 and 3, characterized in that In the interest-aware graph propagation module, lightweight graph convolution is used to perform K I The message is propagated again, and the final representation of users and items in the user interest space is generated through average pooling to fully capture the high-order collaborative information in the user-item bipartite graph. Specifically:
7. The recommendation method according to claims 1, 2, 3, 4, 5 and 6, characterized in that In the cross-view contrastive learning module, the social and sequence information learned from the user-item bipartite graph is used as a self-supervisory signal to optimize the social influence and sequence influence representations learned from the social graph and sequence graph, respectively, thereby promoting the deep dissociation of the two influencing factors. Specifically: Use the attention mechanism to calculate the user's social influence representation and friend preference representation: Use contrastive learning to enhance user social influence and friend preference consistency: Similarly, we optimize the contrastive learning of sequence impact representation:
8. The recommendation method according to claims 1, 2, 3, 4, 5, 6 and 7, characterized in that After executing the multi-channel graph propagation module and the cross-view comparative learning module, the user’s social influence representation can be obtained. and sequence-affected representation Through the adaptive allocation mechanism based on self-attention, the model can dynamically distinguish the degree of influence of different influencing factors on user interests.