Session recommendation method based on time interval perception graph neural network and collaborative enhancement

Abstract

The invention discloses a session recommendation method based on a time interval perception graph neural network and collaborative enhancement, and aims to improve the modeling capability of a recommendation system for time sensitivity, structure discrimination and noise robustness. The method comprises the following four key steps: firstly, constructing a session graph of multiple types of edges containing time interval information based on a session sequence; secondly, fusing time and semantic context features by utilizing a gating attention mechanism, and generating high-quality project embedding; thirdly, by splicing relative time embedding and project embedding, combining GRU and an attention mechanism, extracting dynamic and general preferences of the user, and generating session representation; and finally, a collaborative enhancement mechanism is introduced, information is extracted from similar sessions, and current session representation is enhanced. According to the method, the accuracy and stability of recommendation are improved while the structural modeling precision is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of conversation recommendation systems, and in particular relates to a conversation recommendation method based on time interval perception graph neural network and collaborative enhancement. Background Art

[0002] Session-based recommendation (SBR) is an important technical paradigm for personalized recommendations based on anonymous users' short-term interactions. In real-world scenarios, users often do not log in to the system and express their potential interests through a series of continuous item interactions (such as clicks, browsing, and purchases). This makes traditional recommendation methods that rely on long-term user profiles ineffective. The core challenge of SBR lies in accurately capturing the dynamic evolution of user interests from limited session data and predicting the next item they are likely to interact with.

[0003] Early research focused on three main technical approaches: methods based on sequential pattern mining leveraged Markov properties to model item transition probabilities, but their short-range dependencies limited their ability to capture complex patterns of interest. Recurrent neural network (RNN)-based methods (such as GRU4Rec) learned long-term dependencies through gating mechanisms but neglected critical temporal information in interactive behaviors. While attention-based models (such as NARM and STAMP) were able to identify important items, they were highly sensitive to noisy data. In recent years, graph neural networks (GNNs) have been introduced to the SBR field (such as SR-GNN and GCE-GNN). By constructing conversation graphs to capture high-order transition relationships between items, they have made significant progress. However, existing methods still have three key flaws: First, the use of time information remains at the absolute timestamp level, failing to effectively model the changes in interest intensity reflected by the time intervals between adjacent interactions; second, when constructing the conversation graph, conversations of different time sequences may generate exactly the same topological structure, resulting in the model being unable to distinguish between essentially different user behavior patterns; third, there is a lack of a robust processing mechanism for noise items in the conversation (such as accidental clicks), which will contaminate the overall representation learning through the propagation process of the graph neural network.

[0004] A deeper analysis reveals that existing SBR methods are inherently limited in their temporal sensitivity and structural discrimination. For example, typical GNN methods typically only reflect the frequency of transitions, ignoring the fact that the length of a user's browsing interval from item A to item B may imply a completely different interest transfer intention. Furthermore, when a session contains noisy items that are accidentally clicked, traditional methods are unable to effectively identify them at the single-session level and struggle to leverage cross-session collaborative information for correction. These limitations directly restrict the accuracy and robustness of recommendation systems in real-world scenarios, necessitating the development of novel recommendation frameworks that simultaneously integrate time interval awareness, structural discrimination, and noise suppression. Summary of the Invention

[0005] In terms of technical implementation, this paper first constructs a time-enhanced conversation graph structure, expanding the traditional edge weights that only consider item transfer relationships to dynamic propagation weights that include time interval characteristics. Through the designed time-aware attention mechanism, the model can automatically identify the differences in interest intensity reflected by the time intervals between adjacent interactions, thereby distinguishing user behaviors that are superficially similar but have different temporal patterns. To address the problem of graph structure ambiguity, the innovative relative time embedding technology is introduced to deeply integrate the temporal position information of items in the sequence with their semantic features, effectively resolving the problem of confusion in conversation representations of the same graph structure but different interaction orders.

[0006] To further enhance model robustness, this paper proposes a denoising mechanism based on collaborative enhancement of similar conversations. This mechanism calculates the semantic similarity between conversations, screens out similar conversations with complementary information, and uses gated attention to integrate their pure feature representations into the current conversation learning. This invention enables precise matching of time-aware conversation representations with candidate items.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A conversation recommendation method based on time interval-aware graph neural network and collaborative enhancement includes the following steps:

[0009] Step 1: The core of this step is to convert the conversation sequence into a graph structure rich in temporal information. For a given conversation sequence, it not only records the transfer relationship between items but also accurately captures the time intervals between adjacent interactions. By constructing four types of edges, the model can be used to characterize the temporal characteristics of user behavior.

[0010] In step 1, given a conversation sequence Design a session directed graph G s =(V s ,E s ).in represents the i-th interaction item in the conversation; Indicates the timestamp of the interaction; l is the length of the current session, the session directed graph G s =(V s ,E s ) in E s Defined as a triple (v i ,v j ,Δt ij ), where Δt ij Indicates that the user is from project v i To project v jTime interval, four types of edges should be constructed based on the session: outgoing edge, incoming edge, bidirectional edge and self-loop edge. If (v i ,v j ) appears multiple times, and the average time interval needs to be calculated as the weight representation of the edge, and finally the type embedding vector of each edge is obtained.

[0011] In step 2, the gated attention mechanism is used to dynamically fuse time intervals with item semantic information. The model not only focuses on "which items the user clicks", but also quantifies "how fast the user clicks them". This method uses attention weight allocation to distinguish the semantic relevance of adjacent interactions and enhance the effectiveness of item embedding.

[0012] In step 2, the time interval and the item semantic features are deeply integrated by designing a time-aware graph attention mechanism, and the time interval feature is extracted by using a gated graph attention network. i ,v j ), whose time interval features are obtained through two layers of nonlinear transformation:

[0013]

[0014] ω=tanh(XW ω +b ω )Δt=sigmod(ε⊙W t1 +ω⊙W t2 +b t )

[0015] Among them, ε is the time feature, ω is the context semantic feature, Δt is the fusion of the two feature information, M t =|t i -t j | is the time interval matrix, X is the item feature matrix, W ∈ 、W ω 、W t1 、W t2 is the trainable matrix, b ∈ 、b ω 、b t is the bias term.

[0016] Considering that each node contributes differently to the target node, we need to calculate the node v i For node v j Attention coefficient:

[0017]

[0018] Among them, ⊙ represents element-wise multiplication, represents the edge type embedding vector, Indicates vi In the embedding representation of the l-1th layer, σ represents the activation function, here we use LeakyReLU.

[0019] In order to more conveniently compare the importance of each node, we use the softmax function to perform normalization:

[0020]

[0021] in, It is project v i The neighbor items are finally updated with feature aggregation, and the weighted sum is used to calculate the item v i The embedding representation of is:

[0022]

[0023] Through graph neural networks, the embedding of items in the graph can be aggregated from the features of the item itself and its neighbors. In addition, introducing temporal information and semantic information in the attention weight calculation can further enhance the feature aggregation process.

[0024] Step 3: This step comprehensively models the evolution of user interests by fusing relative time embeddings with item semantic information. The GRU network models the temporal sequence of the sequence, capturing the dynamic changes in user interests. Simultaneously, the attention mechanism is incorporated to automatically identify the historical interactions that are most important to the current recommendation goal. This method also extracts global preference features for the session through average pooling, generating an embedding representing the current session.

[0025] In step 3, the conversation representation learning achieves accurate characterization of user interests through multi-level temporal modeling and attention fusion, which specifically includes the following processes: temporal enhanced item representation generation, based on the item embedding features output in step 2, further fusing relative time information to enhance temporal expression capabilities. For each item v in the conversation i , the feature vector x generated by its graph attention network i and the corresponding relative time embedding t i Perform splicing and obtain time-aware item representation through a nonlinear transformation layer:

[0026] p i =σ(W p [x i ||t i ]+b p )

[0027] Among them, t i Indicates project v i Time interval embedding with the target item, and is a trainable parameter, and || represents a vector concatenation operation. This step effectively distinguishes the differential contributions of the same item to user interest at different time locations. A GRU network is then used to encode the temporally enhanced item representation sequence to capture the dynamic evolution of user interest:

[0028]

[0029] Among them, p′ i represents the hidden state output by the i-th GRU unit, is the GRU network parameter. Through the temporal memory capability of GRU, the model can identify the continuous changes and sudden transfer characteristics of user interests. Considering that the user session as a whole reflects their stable preferences, the average value of all item representations is calculated as the global interest representation:

[0030]

[0031] This global feature complements the dynamic GRU output and together constitutes a complete expression of user interests. A dual-path attention mechanism is designed to integrate dynamic temporal features with global stability features:

[0032]

[0033] in is the weight matrix is the bias vector, and the final conversation representation is the attention-weighted sum of each item representation. This design enables the model to adaptively focus on the historical interaction items that are most indicative of the current recommendation goal:

[0034]

[0035] In step 4, by calculating the similarity between the current session and other sessions, the top-N sessions with collaborative preferences are screened out and their representations are fused in a weighted manner. After collaborative enhancement, the influence of noise is suppressed to obtain the final session embedding representation. Finally, the soft attention mechanism is used to learn and obtain the final recommended items.

[0036] In step 4, robust recommendation is achieved through cross-session collaborative enhancement and hierarchical prediction, specifically including: similar session retrieval and screening, building a session similarity evaluation network in the current training batch, and calculating the target session s through multi-layer perception. c Conversations with other i Deep semantic similarity:

[0037]

[0038] Select the N sessions with the highest similarity to form a similar session set s sim, where N is a preset hyperparameter. Then, attention-weighted fusion is performed on the similar session set to generate a collaborative enhancement representation:

[0039]

[0040] Dynamically fuse the original session representation and the collaborative enhanced representation through a gating mechanism:

[0041] θ=sigmoid(W θ [s c ||s sim ])

[0042] s final =(1-θ)s c +θs sim

[0043] in is a trainable parameter, and || represents vector concatenation.

[0044] Calculate candidate items v i The matching score with the final session representation is used to generate the recommendation probability distribution through the softmax function:

[0045]

[0046] Among them, σ is the softmax function.

[0047] Finally, the cross entropy loss function is used to optimize the model:

[0048]

[0049] The beneficial effects of the conversation recommendation method combining time interval perception and co-graph enhancement provided by the present invention are:

[0050] (1) This paper proposes a novel time interval-aware graph attention mechanism, which dynamically integrates time interval information into the propagation process of graph neural networks and effectively distinguishes behavioral differences in the same sequence but different time patterns.

[0051] (2) This paper designs a multi-level representation learning framework that integrates relative time embedding, GRU dynamic modeling and attention aggregation, which solves the problem of fuzzy representation of different temporal patterns of the same graph structure in traditional methods.

[0052] (3) The present invention innovatively utilizes group information of similar conversations for representation enhancement and suppresses noise interference through cross-conversation knowledge transfer. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the existing technical solutions, the following briefly introduces the drawings required for use in the embodiments. The drawings in the specification are only used to provide a further understanding of the present invention.

[0054] Figure 1 A schematic diagram showing the comparison of different time intervals of the same conversation sequence in the present invention;

[0055] Figure 2 A comparative schematic diagram of generating the same graph structure at different time sequences in the present invention;

[0056] Figure 3 This is a conversation framework diagram with a time interval-aware graph neural network and collaborative enhancement according to the present invention. DETAILED DESCRIPTION

[0057] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0058] The dataset contains multiple sessions S = {s1,s2,…,s n}, given the current session The goal is to recommend the top N items that users are most likely to click from a large set of items. The following is a detailed description of the system process:

[0059] Step 1. The core of this step is to transform the conversation sequence into a graph structure rich in temporal information. For a given conversation sequence, the model not only records the transition relationships between items but also accurately captures the time intervals between adjacent interactions. By constructing four types of edges, the model can be used to characterize the temporal characteristics of user behavior.

[0060] Step 2: This step dynamically integrates time intervals with item semantic information through a gated attention mechanism. The model not only focuses on which items the user clicks, but also quantifies how quickly they click. This method uses attention weighting to distinguish the semantic relevance of adjacent interactions, enhancing the effectiveness of item embeddings.

[0061] Step 3: This step comprehensively models the evolution of user interests by fusing relative time embedding with item semantic information. The GRU network is used to model the temporal sequence and capture the dynamic changes in user interests. At the same time, the attention mechanism is combined to automatically identify the historical interaction items that are most important to the current recommendation target. This method also extracts the global preference features of the session through average pooling to obtain the current session representation embedding.

[0062] Step 4: This step calculates the similarity between the current session and other sessions, screens out the top-N sessions with collaborative preferences, and fuses their representations in a weighted manner. After collaborative enhancement, the influence of noise is suppressed, and the final session embedding representation is obtained. Finally, the soft attention mechanism is used to learn and obtain the final recommended items.

[0063] In step 1, given a conversation sequence Design a session directed graph G s =(V s ,E s ).in represents the i-th interaction item in the conversation; Indicates the timestamp of the interaction; l is the length of the current session, the session directed graph G s =(V s ,E s ) in E s Defined as a triple (v i ,v j ,Δt ij ), where Δt ij =|t i -t j | indicates that the user is from project v i To project v j Time interval, four types of edges are constructed based on the session: outgoing edge e out 、Enter edge e in , bidirectional edge e both and self-loop edge e loop , if (v i ,v j ) appears multiple times, and the average time interval needs to be calculated as the weight representation of the edge, and finally the type embedding vector of each edge is obtained.

[0064] In step 2, based on the traditional graph attention network (GAT), we proposed a time interval-aware attention weight calculation method, which enables the model to dynamically adjust the information transmission intensity between nodes. The specific calculation process is as follows: By designing a time-aware graph attention mechanism, the time interval and the item semantic features are deeply integrated. The gated graph attention network is used to extract the time interval feature. For the edge (v i ,v j ), time interval matrix Defined as M t (i,j)=|t i -t j |. Through nonlinear transformation, temporal features and semantic features are extracted respectively:

[0065]

[0066] ω=tanh(XW ω +b ω )

[0067] Among them, ε is the time feature, ω is the context semantic feature, Δt is the fusion of the two feature information, M t =|t i -t j | is the time interval matrix, X is the item feature matrix, W ∈ 、W ω is the trainable matrix, b ∈ 、b ω is the bias term, and ⊙ represents element-wise multiplication.

[0068] The temporal feature ε and the semantic feature ω are fused to obtain the time-aware attention adjustment factor Δt:

[0069] Δt=sigmod(ε⊙W t1 +ω⊙W t2 +b t )

[0070] Considering that each node contributes differently to the target node, we need to calculate the node v i For node v j Attention coefficient:

[0071]

[0072] Among them, ⊙ represents element-wise multiplication, represents the edge type embedding vector, Indicates v i In the embedding representation of the l-1 layer, σ represents the activation function, and LeakyReLU is used here.

[0073] In order to more conveniently compare the importance of each node, the softmax function is finally used to perform normalization:

[0074]

[0075] in, It is project v i The neighbor items are finally updated with feature aggregation, and the weighted sum is used to calculate the item v i The embedding representation of is:

[0076]

[0077] Through graph neural networks, the embedding of items in the graph can be aggregated from the features of the item itself and its neighbors. In addition, introducing temporal information and semantic information in the attention weight calculation can further enhance the feature aggregation process.

[0078] In step 3, session representation learning achieves accurate characterization of user interests through multi-level temporal modeling and attention fusion, which specifically includes the following processes: temporal enhanced item representation generation, based on the item embedding features output in step 2, further fusing relative time information to enhance temporal expression capabilities, for each item v in the session i , the feature vector x generated by its graph attention network i and the corresponding relative time embedding t i Perform splicing and obtain time-aware item representation through a nonlinear transformation layer:

[0079] p i =σ(W p [x i ||t i ]+b p )

[0080] Among them, t i Indicates project v i Time interval embedding with the target item, and is a trainable parameter, and || represents a vector concatenation operation. This step effectively distinguishes the differential contributions of the same item to user interest at different time locations, resolving the problem that traditional methods cannot distinguish between "early clicks" and "recent clicks." A GRU network is then used to encode the temporally enhanced item representation sequence, capturing the dynamic evolution of user interest:

[0081]

[0082] This formula uses a bidirectional GRU to capture both forward and backward dependencies, p′ i represents the hidden state output by the i-th GRU unit, which encodes the interest evolution path from the beginning of the session to the current position; are the GRU network parameters, which can be learned through end-to-end training. Through the GRU's temporal memory capability, the model is able to identify both continuous changes and sudden shifts in user interests. Considering that the user's overall session reflects their stable preferences, the average of all item representations is calculated as the global interest representation:

[0083]

[0084] This global feature complements the dynamic GRU output and together they form a complete expression of user interest. A dual-path attention mechanism is designed to integrate dynamic temporal features with global stability features:

[0085]

[0086] in is the weight matrix is the bias vector, and the final conversation representation is the attention-weighted sum of each item representation. This design enables the model to adaptively focus on the historical interaction items that are most indicative of the current recommendation goal:

[0087]

[0088] In step 4, robust recommendation is achieved through cross-session collaborative enhancement and hierarchical prediction, specifically including: similar session retrieval and screening. There may be noise problems in a single session, which may affect the accurate capture of user preferences. This noise may not exist in other sessions with similar preferences. Therefore, the current session representation is enhanced by utilizing the representations of other sessions with similar preferences. In the current training batch, a session similarity evaluation network is constructed, and the target session s is calculated through multi-layer perception. c Conversations with other i Deep semantic similarity:

[0089]

[0090] Select the N sessions with the highest similarity to form a similar session set s sim , where N is a preset hyperparameter. Then, attention-weighted fusion is performed on the set of similar sessions to generate a collaborative enhancement representation:

[0091]

[0092] The original session representation and the collaborative enhancement representation are dynamically fused through a gating mechanism, that is, the current session representation and the session representations with similar preferences are aggregated to obtain the final session representation:

[0093] θ=sigmoid(W θ [s c ||s sim ])

[0094] s final =(1-θ)s c +θs sim

[0095] in is a trainable parameter, || represents vector concatenation, s final Represents the final session representation.

[0096] Calculate candidate items v iThe matching score with the final session representation is calculated and normalized by applying the inner product to the recommendation probability of each candidate item:

[0097]

[0098] Among them, σ is the softmax function.

[0099] Finally, the following cross entropy loss function is used to optimize the model:

[0100]

[0101] where y i represents the encoding vector of the real item, Represents the predicted probability.

[0102] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A conversation recommendation method based on time interval-aware graph neural network and collaborative enhancement, comprising the following steps: Step 1: The core of this step is to convert the conversation sequence into a graph structure rich in temporal information. For a given conversation sequence, it not only records the transfer relationship between items but also accurately captures the time intervals between adjacent interactions. By constructing four types of edges, the model can be used to characterize the temporal characteristics of user behavior. Step 2: This step dynamically integrates time intervals and item semantic information through a gated attention mechanism. The model not only focuses on "which items the user clicks" but also quantifies "how quickly the user clicks them." This method uses attention weight allocation to distinguish the semantic relevance of adjacent interactions and enhance the effectiveness of item embedding. Step 3: This step comprehensively models the evolution of user interests by fusing relative time embedding with item semantic information. The GRU network is used to model the temporal sequence and capture the dynamic changes in user interests. At the same time, the attention mechanism is combined to automatically identify the historical interaction items that are most important to the current recommendation target. This method also extracts the global preference features of the session through average pooling to obtain the current session representation embedding. Step 4: This step calculates the similarity between the current session and other sessions, screens out the top-N sessions with collaborative preferences, and fuses their representations in a weighted manner. After collaborative enhancement, the influence of noise is suppressed, and the final session embedding representation is obtained. Finally, the soft attention mechanism is used to learn and obtain the final recommended items.

2. The conversation recommendation method based on time interval perception graph neural network and collaborative enhancement according to claim 1, characterized in that: In step 1, given a conversation sequence Design a session directed graph G s =(V s ,E s ),in represents the i-th interaction item in the conversation; Indicates the timestamp of the interaction; l is the length of the current session, the session directed graph G s =(V s ,E s ) in E s Defined as a triple (v i ,v j ,Δt ij ), where Δt ij Indicates that the user is from project v i To project v j Time interval, four types of edges should be constructed based on the session: outgoing edge, incoming edge, bidirectional edge and self-loop edge. If (v i ,v j ) appears multiple times, and the average time interval needs to be calculated as the weight representation of the edge, and finally the type embedding vector of each edge is obtained.

3. The conversation recommendation method based on time interval perception graph neural network and collaborative enhancement according to claim 1, characterized in that: In step 2, the time interval and the item semantic features are deeply integrated by designing a time-aware graph attention mechanism, and the time interval feature is extracted by using a gated graph attention network. i ,v j ), whose time interval features are obtained through two layers of nonlinear transformation: ω=tanh(XW ω +b ω ) Δt=sigmod(ε⊙W t1 +ω⊙W t2 +b t ) Among them, ε is the time feature, ω is the context semantic feature, Δt is the fusion of the two feature information, M t =|t i -t j | is the time interval matrix, X is the item feature matrix, W ∈ 、W ω 、W t1 、W t2 is the trainable matrix, b ∈ 、b ω 、b t is the bias term, Considering that each node contributes differently to the target node, we need to calculate the node v i For node v j Attention coefficient: Among them, ⊙ represents element-wise multiplication, represents the edge type embedding vector, Indicates v i In the embedding representation of the l-1 layer, σ represents the activation function, here we use LeakyReLU, In order to more conveniently compare the importance of each node, we use the softmax function to perform normalization: in, It is project v i The neighbor items are finally updated with feature aggregation, and the weighted sum is used to calculate the item v i The embedding representation of is: Through graph neural networks, the embedding of items in the graph can be aggregated from the features of the item itself and its neighbors. In addition, introducing temporal information and semantic information in the attention weight calculation can further enhance the feature aggregation process.

4. The conversation recommendation method based on time interval perception graph neural network and collaborative enhancement according to claim 1, characterized in that: In step 3, session representation learning achieves accurate characterization of user interests through multi-level temporal modeling and attention fusion. The process includes the following: temporal enhanced item representation generation, based on the item embedding features output in step 2, further fusing relative time information to enhance temporal expression capabilities, for each item v in the session i , the feature vector x generated by its graph attention network i and the corresponding relative time embedding t i Perform splicing and obtain time-aware item representation through a nonlinear transformation layer: p i =σ(W p [x i ||t i ]+b p ) Among them, t i Indicates project v i Time interval embedding with the target item, and is a trainable parameter, and || represents a vector concatenation operation. This step effectively distinguishes the differential contributions of the same item to user interest at different time locations. Then, a GRU network is used to encode the temporally enhanced item representation sequence to capture the dynamic evolution of user interest: Among them, p′ i represents the hidden state output by the i-th GRU unit, are the GRU network parameters. Through the temporal memory capability of GRU, the model can identify the continuous changes and sudden transfer characteristics of user interests. Considering that the user session as a whole reflects their stable preferences, the average value of all item representations is calculated as the global interest representation: This global feature complements the dynamic GRU output and together constitutes a complete expression of user interests. A dual-path attention mechanism is designed to integrate dynamic temporal features with global stability features: Among them W2, W3, is the weight matrix is the bias vector, and the final session representation is the weighted sum of the attention represented by each item, as follows: This design enables the model to adaptively focus on the historical interaction items that are most indicative for the current recommendation goal.

5. The conversation recommendation method based on time interval perception graph neural network and collaborative enhancement according to claim 1, characterized in that: In step 4, robust recommendation is achieved through cross-session collaborative enhancement and hierarchical prediction, specifically including: similar session retrieval and screening, building a session similarity evaluation network in the current training batch, and calculating the target session s through multi-layer perception. c Conversations with other i Deep semantic similarity: Select the N sessions with the highest similarity to form a similar session set s sim , where N is a preset hyperparameter. Then, attention-weighted fusion is performed on the similar session set to generate a collaborative enhancement representation: Dynamically fuse the original session representation and the collaborative enhanced representation through a gating mechanism: θ=sigmoid(W θ [s c ||s sim ]) s final =(1-θ)s c +θ ssim in is a trainable parameter, || represents vector concatenation, Calculate candidate items v i The matching score with the final session representation is used to generate the recommendation probability distribution through the softmax function: Among them, σ is the softmax function, Finally, the cross entropy loss function is used to optimize the model: where y i represents the encoding vector of the real item, Represents the predicted probability.

Images