Graph-enhanced joint attention method for social time-series recommendation

Abstract

The application discloses a graph-enhanced co-attention method for social temporalized recommendation. Temporalized recommendation is to predict the next interested item according to the behavior sequence of each user. Although many works have dynamically captured the sequential patterns between items, they still face two shortcomings: (1) unable to model the translation patterns between items in a time-aware manner; (2) ignoring the importance of dynamically capturing social influence. The application introduces a new concept of social temporalized recommendation and proposes a graph-enhanced co-attention model (GCARec for short) to dynamically model social influence and capture the translation patterns between items in a time-aware manner. The method firstly adopts a linear graph convolution module to model the long-term preference of users, and then designs time-aware and social-aware attention mechanisms to dynamically capture the sequential patterns and social influence at the same time.

Description

[0001] Technology Space

[0002] The present invention relates to the fields of machine learning and information retrieval, and in particular to methods for using graph neural networks and attention mechanisms for social recommendation and sequential recommendation. Background Art

[0003] In an era of information overload, personalized recommendations have become a crucial component of various commercial applications. They not only deliver personalized information to customers but also significantly increase profits for content providers. Learning users' preferences for target items from their historical behavior is key to personalized recommendation systems. Traditional recommendation systems often model user behavior data statically, ignoring the fact that user interests "drift" over time. In recent years, sequential recommendation methods, such as Markov chain-based methods, recurrent neural networks, convolutional neural networks, and self-attention mechanisms, have garnered widespread attention from both academia and industry due to their high practicality. However, existing sequential recommendation methods primarily focus on user behavior sequences to capture the shifting patterns between items. This can make it difficult to fully tap into short-term user interests, especially for cold-start users.

[0004] With the popularity of social networks, many platforms (such as Douban) encourage users to engage in online entertainment and share experiences with friends. Based on the principles of social influence and homophily, social relationships play a crucial role in providing personalized services to users. This is also known as social recommendation, which leverages social relationships to alleviate two major challenges of collaborative filtering recommendation methods: data sparsity and cold start. In recent years, this field has evolved from early decomposition-based methods to neural social methods and more recently, graph neural network methods. Despite significant progress, existing social recommendation methods discard sequential information about user behavior, resulting in suboptimal recommendation performance.

[0005] To the best of our knowledge, there is very limited work on using social relationships to improve the performance of sequential recommendations. DGRec is a recent work that models dynamic sequential patterns based on long short-term memory networks and uses an attention-coupled graph neural network to capture static social influence. Another work uses an attention recurrent neural network to model dynamic user interests and employs an attention mechanism to capture social influence. While both approaches have achieved promising results, they still have the following shortcomings:

[0006] (1) At each time step, these RNN-based recommendation methods heavily rely on the hidden state at the previous time step and the current state. In this case, they cannot learn the shift patterns between non-adjacent items and are difficult to implement parallel computing. In addition, these methods usually assume that the user's behavior record is ordered, ignoring the time interval between the user's recent behavior and the target behavior. Intuitively, the closer the interaction time of an accessed item to the user's current time, the greater the influence of the item on the user's future decision.

[0007] (2) These methods ignore the importance of dynamically capturing social influence and fail to explicitly model the dynamic influence between users and friends. To illustrate this defect, Figure 1 A small example in the movie scene is shown. At time t1, Lily is looking for a science fiction movie, while his friends A and B have completely opposite tastes. Considering the current situation, it is more appropriate to recommend Lily to watch the movie "Avatar" based on the recommendation of friend A. At time t2 (such as Saturday night), she wants to watch a romantic movie with her boyfriend. In this case, friend B should have a greater influence on her, because the friend is obsessed with movies on the theme of love. SUMMARY

[0008] To address the above defects or problems, a new recommendation concept is disclosed: social-temporal recommendation, to study the joint effect of dynamic sequential patterns and dynamic social influence. In addition, to solve this task, a graph-enhanced joint attention model (GCARec) is proposed. The proposed method decomposes the social-temporal task into two steps:

[0009] First, a linear graph convolution module is used to model the user's long-term preferences.

[0010] Then, according to the shift patterns between items and the social relationships between users, a joint attention module is designed to dynamically capture sequential patterns and social influence. In particular, a time-aware attention mechanism is introduced to capture the complex dependency relationship between items. On this basis, a social-aware attention mechanism is further constructed to measure the influence of different friends on the target user's decision at the current time. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description.

[0012] Figure 1 is a schematic diagram of a recommendation scene of an embodiment of the present application;

[0013] Figure 22 is a schematic diagram of the GCARec model structure according to an embodiment of the present invention;

[0014] Figure 3 This is a comparison chart of social time-series recommendation and other recommendation tasks according to an embodiment of the present invention;

[0015] Figure 4 2 is a graph comparing the results of GCARec under different evaluation indicators according to an embodiment of the present invention;

[0016] Figure 5 This is a comparison chart of results under different recommendation list lengths provided by an embodiment of the present invention;

[0017] Figure 6 This is a result comparison chart under different user groups provided by an embodiment of the present invention;

[0018] Specific implementation

[0019] This paper first provides a formalization of the problem to be solved in this work. Then, it elaborates on how GCARec models users' long-term preferences and short-term interests. Finally, it discusses the space complexity and time complexity.

[0020] 1. Problem Definition:

[0021] Definition 1: Social temporal recommendation task

[0022] Input: User Set Item Set Each user Behavior sequence and the user's friends collection

[0023] Output: from the user u has not visited before but may interact with it in the future Export the top N item lists in .

[0024] exist Figure 3 In

[15] , the inventors provide a visual comparison of social temporal recommendation and other recommendation tasks.

[0025] In this invention, bold letters are used to represent vectors and matrices, such as e and W. Capital calligraphic letters are used to represent sets, such as For a better understanding of the present invention, Table 1 summarizes the mathematical symbols and their definitions used in the present invention.

[0026] Table 1 Symbols

[0027]

[0028] 2. Overall architecture of the model:

[0029] Overall architecture of the model: Figure 2The overall architecture of the GCARec model is shown. The model consists of four core components: representation initialization, a linear graph convolution module, a joint attention module, and model prediction. The following text explains each component in detail.

[0030] 2.1 Characterization Initialization:

[0031] Following the mainstream serialized recommendation model, the present invention uses the one-hot encoding technique to initialize each user / item into a vectorized representation (representation). Specifically, the present invention uses and Initialize user u and item i respectively, where D represents the dimension of the vector. The present invention also initializes a vector for each item Used to capture the serialization mode and additionally initialize a vector for each user Used to capture social influence. Based on this, the representation of all items and users can be obtained as follows:

[0032]

[0033] Where [·] represents the concatenation operation of vectors.

[0034] 2.2 Linear Graph Convolution Module:

[0035] In previous sequential recommendation methods, such as FPMC, SASRec, and HGN, the initialized user representation (i.e. ) is directly used to capture users' long-term preferences. Although this approach is simple and easy to think of, it is difficult to obtain satisfactory recommendation results. The key reason is that these methods only explore the observed direct interaction relationships, but cannot capture the high-order connectivity information between users and items. In industrial scenarios, the main obstacle of a recommendation system is the problem of data sparsity. Explicitly capturing collaborative signals in the user-item interaction graph is particularly important for obtaining high-quality node representations, especially for inactive users. Inspired by LightGCN, the inventors corrected the initialized representation matrix E through a linear graph convolution module. Specifically, the module consists of two parts: neighbor aggregation and representation readout.

[0036] (1) Neighbor aggregation: To explicitly model the collaborative signals in the user-item interaction graph, recommendation methods based on graph neural networks usually use neighbor aggregation to collect features from neighboring users. Taking the user node as an example, the kth layer of a graph propagation module can be abstracted as:

[0037]

[0038] where f propagateis a propagation function. By using different propagation functions, the above formula can realize different graph neural network models, such as weight summation in GCN, bilinear interaction in NGCF, and self-attention mechanism in KGAT. Recent studies have shown that f propagate The two common operations in [1] — nonlinear activation and feature transformation — will reduce the performance of the model in the recommendation scenario. In view of this, the present invention adopts the simplest operation — linear aggregator, and then formula (2) can be rewritten as:

[0039]

[0040] in

[0041] represents the set of users who have visited item i.

[0042] (2) Representation readout: After stacking multiple layers of linear graph convolution, the representation of each user and item in different convolutional layers can be obtained. To alleviate the over-smoothing problem, the present invention adopts an average pooling strategy to generate the final representation of users and items, which is calculated as follows:

[0043]

[0044] 2.3 Joint Attention Module:

[0045] How to capture short-term user interests is the key to improving the performance of the recommendation system. To achieve this goal, the present invention explores dynamic patterns from two levels, namely serialization and socialization. Still using the previous example, at time t1, compared with the movie "Flipped", the movies "WALL·E" and "Transformers" contribute more to predicting the user's current interest. In addition, the closer the item a user visits is to his current time, the greater the impact of the item on the user's future decision. Secondly, dynamic modeling of social influence is also crucial to capturing users' short-term interests. For example, at time t1, friend A has a greater influence on Lily's decision, while at time t2 the situation is the opposite. In view of this, the present invention designs a joint attention module to model users' short-term interests, including a time-aware attention mechanism and a social-aware attention mechanism, which are described in detail below.

[0046] (1) Time-aware attention mechanism: For each user, the present invention selects L recent items as the input of the component and takes the item i to be accessed at the next moment as the prediction target, where Constitute an L-order Markov chain. Then, the interest of user u at time t can be abstracted as:

[0047]

[0048] where fagg is an aggregation function that aims to integrate the L-order Markov chain into a unified vector. Common maximum pooling or average pooling operations can be directly applied here. However, these two strategies fail to distinguish the importance of each candidate item in predicting the next item. For example, the viewing history of user u is The viewing history of user v is Although the order of actions of the two users is the same, the sequence The short-term interest of user u is greater than The contribution of user v is greater. This is mainly because user u watched these movies in one day, while user v generated these behaviors within a week. However, most sequential recommendation methods (such as HGN, SASRec, DGRec) usually only focus on the order of occurrence of behaviors in the sequence and ignore the time interval information between behaviors. To solve this problem, the present invention designs a time-aware attention mechanism that can dynamically determine the contribution of each item in the L-order Markov chain. Given the current timeline of user u The time interval feature can be introduced based on the following formula:

[0049]

[0050] in, φ(·) is a tanh activation function. The time interval feature δ(u,tl,t) encodes the absolute time information between time tl and time t. On this basis, the attention vector can be formalized as:

[0051]

[0052] in, and A learnable weight matrix for distilling high-quality features. To reduce model parameters and obtain personalized patterns, the present invention uses As the context vector. Afterwards, by h(u,l,t) and Applying the softmax function above yields the weight value α(u,l,t):

[0053]

[0054] After obtaining the attention weight value, formula (5) can be rewritten as:

[0055]

[0056] (2) Socially Aware Attention Mechanism: Reasonable modeling of social influence is the key to social temporal recommendation, because dynamically modeling the social relationships between different users helps to explain "Why does a user like this particular item at a certain moment?". However, current recommendation methods ignore social relationships (such as ATM, SR-GNN and SASRec), or capture social influence in a time-independent way (such as DiffNet and EAGCN). In fact, accurately modeling the social influence of users is not easy. For example, whether Lily watches the movie "Avata", she may refer more to the recommendation of friend A; this is because friend A has just watched the movie and has also watched many other science fiction movies, such as "Inception" and "Transformer". At time t2, friend A's interest has a relatively small impact on Lily. The reason is that friend A has very few movies related to love themes in his historical viewing records, so it has little influence on Lily's decision. The present invention designs a socially aware attention mechanism to distinguish the influence of each friend on the target user's next decision at different times. The calculation method of the attention vector is as follows:

[0057] h(u,v,t)=φ(z u (t)W3+q v W4+b′) (10)

[0058] in and is a trainable model parameter. Same as formula (8), here we also use As the context vector. By using the softmax function, and in h(u,v,t) and Normalizing the above equations yields the weight α(u,v,t):

[0059]

[0060] After obtaining the weight value of social perception, the influence of each friend can be calculated based on the following formula:

[0061]

[0062] 2.4 Model Prediction:

[0063] After executing the linear graph convolution module and the joint attention module, the user's long-term preference can be obtained Short-term interest with serialization pattern u (t) and short-term interests y containing dynamic social influences u (t). Finally, user u’s preference for item i can be calculated based on the following formula:

[0064]

[0065] where μ i for capturing the popularity of items i.

[0066] 3. Model optimization:

[0067] Given a user u, a friend and a recent behavior sequence The probability of item i being visited at time t can be derived as:

[0068]

[0069] where σ(x) = 1 / (1 + e -x ). Accordingly, the present application optimizes GCARec based on the following objective function:

[0070]

[0071] where λ is the regularization coefficient and Θ is the set of model parameters. The present application employs a strategy of random sampling to select negative samples j from the unobserved data .

[0072] 3.1 Model analysis:

[0073] (1) Space complexity: The model parameters of the GCARec method proposed by the present application are derived from two parts: Θ1 = {E, V, Q, μ} and Θ2 = {ω σ , W1, W2, W3, W4, b, b'}. The former is proportional to the number of users and items (i.e., O(D ). The latter is completely independent of the number of users and items (i.e., 4D 2 + 3D). Through analysis, the model size of GCARec is comparable to the most advanced HGN (i.e., 4D ) and is lighter than the pioneer model FPMC (i.e., 4D ). For example, on the Epinions dataset (12.7K users and 209.7K items), when D is set to 64, FPMC and HGN use 40.12M and 27.22M parameters, respectively, while GCARec uses 28.02M parameters.

[0074] (2) Time complexity: For the long-term preference module of users, the time complexity of calculating vectors and is O(D ), where represents the number of all user behavior records. For the time-aware attention mechanism, the time complexity of calculating h(u, l, t) is O(D 2Since the denominator of the softmax function in formula (8) needs to traverse the L-order Markov chain, the time complexity of calculating an α(u,l,t) is O(D 3 L). For all users, the time complexity of this part is Similarly, the time complexity of the attention mechanism for social perception is in represents the average number of social relationships a user has. In summary, the total time complexity of GCARec is

[0075] Empirical results show that GCARec significantly outperforms mainstream recommendation methods. Furthermore, to further investigate the effectiveness of each module of this method, we conducted ablation experiments to demonstrate that GCARec achieves the best performance compared to other methods. This demonstrates the complementary performance improvements between different modules and further validates the rationality and effectiveness of our proposed method.

[0076] 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.

Claims

1. A graph-enhanced joint attention method for social temporal recommendation, characterized by: The steps include: (1) Representation initialization: Use the one-hot encoding technique to initialize each user and item into a vectorized representation. For each user, an additional vector is introduced to capture social influence, and for each item, an additional vector is introduced to capture serialization patterns. These are used for subsequent modeling of social influence and serialization preference, respectively. (2) Linear graph convolution: Based on the user-item interaction graph, multi-layer linear graph convolution operations are performed to perform neighbor aggregation, explicitly modeling the high-order collaborative relationship between users and items, thereby generating a long-term preference representation of users; (3) Joint attention: It includes the following two submodules: time-aware attention mechanism and social-aware attention mechanism; among them, the time-aware attention mechanism can dynamically adjust the contribution of each recently visited item to the short-term interest modeling, including the following steps: First, according to the current timestamp T of user u, t u Introducing time interval features: Then, calculate the attention vector: After that, in h(u,l,t) and Apply the softmax function to get the weight value: Finally, we can get the interest of user u at time t: The social perception attention mechanism can dynamically adjust the importance of each friend for predicting the user's next behavior at different times, including the following steps: First, calculate the attention vector: h(u,v,t)=φ(z u (t)W3+q v W4+b′); then, the softmax function is used to calculate the h(u,v,t) and Normalize the weights to get the values: Finally, calculate the influence of friends: (4) Model prediction: The user’s long-term preferences obtained through linear graph convolution operations are combined with the short-term interests obtained through the joint attention mechanism that take into account dynamic sequence patterns and social influences to predict the user’s preferences for candidate items.

2. The graph-enhanced joint attention method for social temporal recommendation according to claim 1 is characterized in that: The graph convolution module uses the linear propagation rule to stack multiple graph convolution layers and generates a user's long-term preference representation through average pooling. It includes the following steps: (1) Use the unique hot encoding technique to initialize each user or item into a vectorized representation, and additionally initialize a vector for each item Used to capture the serialization mode and initialize a vector for each user To capture social influence, the representation of all items and users can be expressed as: (2) Using the neighbor aggregation operation to collect features from neighboring users, we explicitly model the collaborative signals in the user-item interaction graph. The representation of users and items at the kth layer of the graph propagation module can be expressed as: (3) Use the average pooling strategy to generate the final representation of users and items:

3. The graph-enhanced joint attention method for social temporal recommendation according to claim 1 or 2, characterized in that: The user's preference score for an item is calculated by fusing long-term preference, time-aware short-term interest, social-aware short-term interest, and the target item representation, and also takes into account the popularity of the target item:

4. The graph-enhanced joint attention method for social temporal recommendation according to claim 1 or 3, characterized in that: The method uses a binary cross entropy loss function based on random negative sampling for optimization and updates the model parameters through the stochastic gradient descent method:

Images