A Multi-Behavior Sequence Recommendation Method and System Based on Graph Neural Networks

Abstract

This invention discloses a multi-behavior sequence recommendation method and system based on graph neural networks. The recommendation method includes acquiring user multi-behavior sequences; inputting the user multi-behavior sequences into a pre-constructed global item graph to obtain a user interaction graph; and using a pre-constructed graph-perceptron mixer model to analyze and mine the user interaction graph to obtain recommendation results. The graph-perceptron mixer model combines the MLP-Mixer architecture and graph neural networks (GNNs). By integrating information from the global item graph and utilizing the interest mining capabilities of the MLP-Mixer, it achieves a deep understanding and representation of user behavior sequences, improving the accuracy of the recommendation method.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and specifically to a multi-behavior sequence recommendation method and system based on graph neural networks. Background Technology

[0002] In the field of recommender systems, multi-behavior sequence recommendation models aim to comprehensively analyze user behavior patterns across different types of interactions (such as clicks, likes, and purchases) to optimize recommendation performance. In recent years, Graph Neural Networks (GNNs) have been widely applied in multi-behavior sequence recommendation due to their superior modeling capabilities and ability to effectively uncover complex behavioral patterns in user sequences. However, traditional GNNs propagate information in the interaction graph through multi-layered local message passing mechanisms, but they are insufficient at modeling collaborative signals between sequences and have significant limitations in handling long-distance dependencies and avoiding overcompression of information. While global attention mechanisms can alleviate these problems, their model complexity is quadratic, resulting in high computational costs, which poses a significant challenge to model efficiency, especially given the ever-increasing number of user behavior sequences. Summary of the Invention

[0003] To address the problems of the existing technologies, this invention provides a multi-behavior sequence recommendation method and system based on graph neural networks. First, a global item graph is constructed to enhance the model's ability to model inter-sequence collaborative signals. Then, a perceptron-mixer architecture is combined with a graph neural network to obtain a graph-perceptron-mixer model that fully mines user interests. This effectively captures the global dependencies of user behavior, significantly improving the time and space efficiency of the model's recommendation. Its complexity is linearly related to the number of user interaction behaviors. The technical solution is as follows:

[0004] Firstly, a multi-behavior sequence recommendation method based on graph neural networks is provided, including the following steps:

[0005] Obtain user multi-behavior sequences;

[0006] Input the user's multi-behavior sequence into a pre-constructed global item graph to obtain the user interaction graph;

[0007] By using a pre-built graph-perceptron hybrid model, user interaction graphs are analyzed and mined to obtain recommendation results.

[0008] In some implementations, the pre-constructed global item graph includes:

[0009] The nodes in the graph represent the set of item nodes that appear in the interaction sequence;

[0010] In the graph, any two item nodes v are connected by an edge. i and v j Weight w of the edges betweenij It is the sum of the number of times the two nodes are connected in all sequences at order k = [1,2,…,n], where k represents the adjacent interval of the items;

[0011]

[0012] w ij Represents two item nodes v i and v j The weight of the edge between them, w′ ij Indicates the relationship with w ij Normalized weights; loc(v i ) represents item v i In the sequence, δ represents the number of times two items differ by k in the sequence, U is the user set, and I... u It is the interaction sequence of user u;

[0013] The construction process of the global item graph also includes: for edges with weights greater than a preset threshold, their weights are reset to 0.

[0014] In some implementations, the data processing steps of the graph-perceptron hybrid model include:

[0015] Interest extraction steps: Divide the user interaction graph into overlapping interest subgraphs;

[0016] Interest embedding steps: Transform the interest subgraph into an initial interest representation;

[0017] Interest mining steps: For the initial interest representation, the final interest representation is generated by modeling each interest information and the relationship between interests through multiple residual networks;

[0018] User embedding steps: For the final interest representation, a user representation is obtained through global average pooling and connected to a fully connected layer for the next item prediction.

[0019] In some implementations, the interest extraction step is implemented using an improved METIS method, which includes:

[0020] Using METIS, the user interaction graph G is divided into P non-overlapping tiles, {v1,...,v...} p}, and v1∪...∪v p =G,

[0021] Extend the tiles to their one-hop neighbors to preserve information about the links between tiles and utilize all graph edges.

[0022] In some implementations, the interest embedding step, which transforms the interest subgraph into an initial interest representation, includes:

[0023] (51) For each interest subgraph, for the original nodes and edge Perform linear embedding and linear projection onto d-dimensional hidden features:

[0024]

[0025]

[0026] in, u 0 ∈R d v 0 ∈R d u 0 v 0 These are learnable parameters;

[0027] (52) Graph update is performed using graph convolutional layers with MP-GNN:

[0028] Nodes and edges are represented by each tile. Apply MP-GNN for updates;

[0029]

[0030]

[0031] in, l is the layer index, p is the tile index, i and j represent nodes, N(i) is the neighborhood of node i; a function f has learnable parameters. node f edge ; These are the average representations of the tile nodes and tile edges, respectively.

[0032] g patch-node g patch-edge It is an MLP-based function that operates on and

[0033] For each node and edge that covers multiple tiles due to tile overlap, update the node / edge representation by averaging the representations in the overlapping tiles:

[0034]

[0035]

[0036] Where {k|i∈v k},{k|ij∈ε k Let} be the set of all graph tiles that cover node i and edge ij;

[0037] (53) G p All node vectors in the array are average pooled, so that... And apply a small MLP to obtain fixed-size tile embeddings.

[0038] In some implementations, the interest embedding step of converting the interest subgraph into an initial interest representation further includes adding location encoding information to nodes and tiles, wherein the location encoding includes:

[0039] (61) Node Location Encoding PE

[0040] Introducing p i ∈R K and a learnable matrix T 0 ∈R d×K Enhance the information of the input node features.

[0041]

[0042] (62) Patch location code PE

[0043] Add tile location encoding to the tile embedding of different tiles Obtain new tile embeddings, including:

[0044] (621) Through the coarsened adjacency matrix A of the interest subgraph P ∈R P×P To capture relative position information: in, It is a standard graph cutting operator that computes the interest subgraph v. i v j The number of connecting edges between them;

[0045] (622) Based on coarsened adjacency matrix A P Get the position code of the i-th tile. Patch-based location encoding Update the embedded representation of the tile Where, x i It is an embedded representation of a tile;

[0046] In some implementations, the interest mining step includes: generating a final interest representation by modeling each interest information and the relationships between interests using multiple residual networks for the initial interest representation, including: processing the interest embedding set using an MLP-Mixer: X∈R P×d Embedding collections for interests

[0047] The steps of interest discovery are represented as follows:

[0048] U=X+(W2δ(W1LayerNorm(X)))∈R P×d ;

[0049] Y = U + (W4δ(W3LayerNorm(U)) T )) T ∈R P×d ;

[0050] Where δ represents the GELU nonlinear activation function, LayerNorm(·) represents the layer normalization operation, and the matrix... Where, d s and d c These are the adjustable hidden layer widths in token-mixing and channel-mixing MLPs, respectively.

[0051] In some implementations, the user embedding step includes: obtaining a user representation from the final interest representation through global average pooling, and connecting it to a fully connected layer for the next item prediction, comprising:

[0052] Average pooling steps:

[0053] Among them, h G ∈R d m p It is a binary variable, with a value of 1 for non-empty blocks and a value of 0 for empty blocks;

[0054] The prediction result u is obtained based on the fully connected layer MLP.

[0055] In some implementations, the interest extraction step further includes:

[0056] For the original graph G = (v, ε), a perturbation graph G′ = (v, ε′) is created, where ε′ is obtained by randomly deleting a small portion of the edges from the original edge set ε;

[0057] An improved METIS algorithm is applied to G′ to obtain different item node partitions {v1,...,v p}

[0058] Secondly, a multi-behavior sequence recommendation system based on graph neural networks is provided, including:

[0059] User multi-behavior sequence acquisition unit, used to acquire user multi-behavior sequences;

[0060] The user interaction graph acquisition unit is used to input multiple user behavior sequences into a pre-constructed global item graph to acquire the user interaction graph.

[0061] A multi-behavior sequence recommendation analysis unit is used to analyze and mine user interaction graphs using a pre-built graph-perceptron hybrid model to obtain recommendation results. The graph-perceptron hybrid model includes:

[0062] The interest extraction module divides the user interaction graph into overlapping interest subgraphs.

[0063] The interest embedding module transforms the interest subgraph into an initial interest representation, which is then input into the interest mining module.

[0064] The interest mining module generates the final interest representation by modeling each interest information and the relationship between interests through multiple residual networks, based on the initial interest representation.

[0065] The user embedding module obtains the user representation by global average pooling based on the final interest representation, and then connects it to the fully connected layer for the next item prediction.

[0066] The present invention provides a multi-behavior sequence recommendation method and system based on graph neural networks, which has the following advantages: The present invention constructs a global graph through the collaborative relationships between items, thereby enhancing the representation of user interaction subgraphs; the graph-perceptron mixer model GraphMLP-Mixer of the present invention can effectively capture the global dependencies of user behavior while alleviating the problem of information over-compression; secondly, its time and space efficiency is significantly improved, and its complexity is linearly related to the number of user interaction behaviors, which is superior to existing multi-behavior sequence recommendation models based on GNNs. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the multi-behavior sequence recommendation method based on graph neural networks proposed in this application;

[0068] Figure 2 This is a schematic diagram of the GraphMLP-Mixer model;

[0069] Figure 3 This is a flowchart of the data processing steps for the GraphMLP-Mixer model;

[0070] Figure 4 This is a schematic diagram of the structure of the multi-behavior sequence recommendation system based on graph neural networks in this application. Detailed Implementation

[0071] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0072] See Figure 1 This application provides a multi-behavior sequence recommendation method based on graph neural networks, which includes the following steps:

[0073] Step 1: Obtain the user's multi-behavior sequence;

[0074] Step 2: Input the user's multi-behavior sequence into the pre-constructed global item graph to obtain the user interaction graph;

[0075] Step 3: Utilize the pre-built graph-perceptron hybrid model to analyze and mine the user interaction graph to obtain recommendation results.

[0076] Furthermore, in step 2 above, the pre-constructed global item map includes:

[0077] The nodes in the graph represent the set of item nodes that appear in the interaction sequence;

[0078] In the graph, any two item nodes v are connected by an edge. i and v j Weight w of the edges between ij It is the sum of the number of times the two nodes are connected in all sequences at order k = [1,2,…,n], where k represents the adjacent interval of the items;

[0079]

[0080] w ij Represents two item nodes v i and v j The weight of the edge between them, w′ ij Indicates the relationship with w ij Normalized weights; loc(v i ) represents item v i In the sequence, δ represents the number of times two items differ by k in the sequence, U is the user set, and I... u It is the interaction sequence of user u;

[0081] For edges whose weights are greater than a preset threshold, their weights are reset to 0.

[0082] In this application, a user interaction graph G is formed by collecting the item nodes that the user has interacted with and the edges between them, based on a pre-constructed global item graph. Let G = (V, ε) represent the subgraph of each user interaction sequence, or the user's interest graph, where V is the set of item nodes and ε is the set of edges. The graph has N = |V| nodes and E = |ε| edges.

[0083] The connectivity of a graph is determined by the adjacency matrix A∈R N×N The feature representation of node i is h. i The feature representation of the edge between nodes i and j is e. ij Let {V1,...,V} p} represents the partitions of the interest subgraph, P is the predefined number of partitions, and G is the number of partitions.i =(V i ,ε i ) is an induced subgraph of G containing all nodes V. i and all endpoints belong to V i The edge. Let h G It is a graph-level representation, y G It is a graph-level target.

[0084] Further, see Figure 2 and Figure 3 In step 2 above, the data processing steps for the graph-perceptron hybrid model include:

[0085] Step 21, Interest Extraction Step: Divide the user interaction graph into overlapping interest subgraphs;

[0086] Step 22, Interest Embedding Step: Convert the interest subgraph into an initial interest representation;

[0087] Step 23, Interest Mining Step: For the initial interest representation, the final interest representation is generated by modeling each interest information and the relationship between interests through multiple residual networks;

[0088] Step 24, User Embedding Step: For the final interest representation, the user representation is obtained through global average pooling and connected to the fully connected layer for the next item prediction.

[0089] In this embodiment, the graph-perceptron mixer model combines an MLP-Mixer architecture and a graph neural network (GNN). The interest extraction step segments the user interaction graph into overlapping interest subgraphs, and the interest embedding step transforms these subgraphs into initial interest representations. These representations are then fed into an interest mining step to generate final representations. User representations are obtained through a global average pooling layer and connected to a fully connected layer for prediction. By integrating information from the global item graph and leveraging the interest mining capabilities of the MLP-Mixer, a deep understanding and representation of user behavior sequences is achieved.

[0090] Furthermore, step 21 above, the interest extraction step, is implemented based on an improved method of METIS. This improved method of METIS, i.e., the interest extraction step, includes:

[0091] Step 211: Use METIS to divide the user interaction graph G into P non-overlapping tiles, {v1,...,v...} p}, and v1∪...∪v p =G,

[0092] Step 212, extend the pieces to their one-hop neighbors to preserve information about the links between pieces and utilize all graph edges. V i ←V i∪{N1(j)|j∈V i}, where Nk(j) defines the k-hop neighbors of node j.

[0093] METIS is a highly efficient graph clustering algorithm renowned for its superior accuracy and efficiency. This algorithm can segment a graph into a predetermined number of clusters, ensuring that the connection density within a cluster is much higher than the connection density between clusters, thus effectively capturing the subgraph structure within the graph. However, a limitation of METIS is that it can only identify non-overlapping clusters. In the context of recommender systems, for example, a user's purchase of a laptop might simultaneously reflect the user's learning and entertainment preferences. Extracting only non-overlapping subgraphs could lead to the loss of important information, thus affecting the accuracy of predictions. To overcome this limitation and utilize all edges in the graph, this application improves the METIS algorithm to generate overlapping subgraphs. This method preserves the integrity of the user's diverse interests, thereby improving the predictive performance of the recommender system.

[0094] Specifically, if the source node and target node of an edge are not in the same subgraph, in this embodiment, the two nodes are assigned to two subgraphs simultaneously. This approach is equivalent to expanding the graph into blocks to include one-hop neighbors of all nodes within each block.

[0095] Furthermore, considering the strong overfitting tendency of the MLP-Mixer architecture, this application introduces a random perturbation strategy in the METIS graph partitioning algorithm to mitigate this problem. Specifically, in the interest extraction step described above, step 211 further includes:

[0096] Step 21101: Create a perturbation graph G′ = (v, ε′) for the original graph G = (v, ε), where ε′ is obtained by randomly deleting a small portion of the edges in the original edge set ε;

[0097] Step 21102: Apply the improved METIS algorithm to G′ to obtain different item node partitions {v1,...,v p}

[0098] Specifically, this embodiment first defines an original graph G = (v, ε), and then creates a perturbation graph G′ = (v, ε′), where ε′ is obtained by randomly deleting a small subset of edges from the original edge set ε. In each iteration, this embodiment applies the improved METIS algorithm to G′ to obtain different item node partitions {v1,...,v...}. p These partitions are generated based on the structural characteristics of the perturbation graph, thus introducing a degree of randomness that helps break the local optima that the model might get stuck in. Then, this application extracts the graph patches {G1,...,G...}. p}, where G i=(V i ,ε i ) is an induced subgraph of G, not a modified G′, so that this application can generate different graph tiles in each iteration, which retains all the node and edge information of the original graph.

[0099] In this way, this application generates a different set of tiles in each iteration, introducing randomness while preserving the original graph structure. This strategy not only increases the model's generalization ability but also improves its robustness to unseen data. The random perturbation strategy effectively breaks specific patterns in the data that the model might over-rely on, thereby helping to mitigate overfitting and enabling the model to maintain stable and efficient performance in a wider range of real-world applications.

[0100] Furthermore, step 22 above, the interest embedding step: converting the interest subgraph into an initial interest representation, includes:

[0101] (51) For each interest subgraph, for the original nodes and edge Perform linear embedding and linear projection onto d-dimensional hidden features:

[0102]

[0103]

[0104] in, u 0 ∈R d v 0 ∈R d u 0 v 0 These are learnable parameters;

[0105] (52) Graph update is performed using graph convolutional layers with MP-GNN:

[0106] Nodes and edges are represented by each tile. Apply MP-GNN for updates;

[0107]

[0108]

[0109] in, l is the layer index, p is the tile index, i and j represent nodes, N(i) is the neighborhood of node i; a function f has learnable parameters. node f edge function f node f edge(With learnable parameters) Define any arbitrary MP-GNN architecture.

[0110] These are the average representations of the tile nodes and tile edges, respectively.

[0111] g patch-node g patch-edge It is an MLP-based function that operates on and

[0112] For each node and edge that covers multiple tiles due to tile overlap, update the node / edge representation by averaging the representations in the overlapping tiles:

[0113]

[0114]

[0115] Where {k|i∈v k},{k|ij∈ε k Let} be the set of all graph tiles that cover node i and edge ij;

[0116] (53) G p All node vectors in the array are average pooled, so that... And apply a small MLP to obtain fixed-size tile embeddings.

[0117] Furthermore, while graph modeling of user sequences strengthens the relationships between multiple user behavior sequences, it easily overlooks the temporal information of each sequence. However, for general graphs, this ordering of nodes and tiles does not exist. This lack of location information reduces the model's ability to express the temporal nature of user behavior. In this embodiment, two explicit location codes (PEs) are used: an absolute PE for item nodes and a relative PE for interest tiles. That is, step 22, the interest embedding step, further includes: adding location coding information to nodes and tiles, wherein the location coding includes:

[0118] (61) Node Location Encoding PE

[0119] In step (51), p is introduced. i ∈R K and a learnable matrix T 0 ∈R d×K Enhance the information of the input node features.

[0120]

[0121] (62) Patch location code PE

[0122] After step (53), add tile location encoding for tile embedding of different tiles. Obtain new tile embeddings, including:

[0123] (621) Through the coarsened adjacency matrix A of the interest subgraph P ∈R P×P To capture relative position information: in, It is a standard graph cutting operator that computes the interest subgraph v. i v j The number of connecting edges between them;

[0124] The relative position information between tiles can be obtained from the adjacency matrix A∈R of the original graph. N×N METIS extracts interest subgraphs {v1,...,v} p The calculation yielded the result.

[0125] (622) Based on coarsened adjacency matrix A P Get the position code of the i-th tile. Patch-based location encoding Update the embedded representation of the tile Where, x i It is an embedded representation of a tile;

[0126] This application extracts location codes at the tile level. Similar to the item node level, this will be injected (after linear transformation) into the first layer of the interest mining step:

[0127]

[0128] Where, x i It is an embedded representation of a tile.

[0129] Furthermore, in step 23 above, the interest mining step: for the initial interest representation, the final interest representation is generated by modeling each interest information and the relationship between interests through multiple residual networks, including:

[0130] The MLP-Mixer is introduced into graph neural networks, and is used to process the interest embedding set: X∈R P×d Embedding collections for interests

[0131] The steps of interest discovery are represented as follows:

[0132] U=X+(W2δ(W1LayerNorm(X)))∈R P×d ;

[0133] Y = U + (W4δ(W3LayerNorm(U)) T )) T ∈R P×d ;

[0134] Where δ represents the GELU nonlinear activation function, LayerNorm(·) represents the layer normalization operation, and the matrix... Where, d s and d c These are the adjustable hidden layer widths in token mixing (within an interest) and channel mixing (between interests) MLPs.

[0135] In this embodiment, the interest mining step combines a GNN with an MLP-Mixer network, employing iterative channel and token mixing steps to achieve information fusion within and between interests. This method significantly reduces computational complexity while maintaining model performance, enabling more efficient mining and utilization of user interest information.

[0136] Furthermore, in step 24 above, the user embedding step involves obtaining a user representation from the final interest representation through global average pooling, and then connecting it to a fully connected layer for the next item prediction, including:

[0137] Step 241, Average Pooling Step:

[0138] Among them, h G ∈R d m p It is a binary variable, with a value of 1 for non-empty regions and a value of 0 for empty regions (since the size of the graph can vary, small graphs may produce empty regions);

[0139] Step 242: Obtain the prediction result u based on the fully connected layer MLP.

[0140] In the training process of the graph-perceptron hybrid model, maximizing Q(u) is used as the optimization objective:

[0141] Where v t ∈v L+1 This represents the item whose interaction with the target behavior is to be predicted.

[0142] In constructing the deep learning model for the recommender system, this embodiment introduces an interest mining step, specifically designed to extract user interest representations. To derive a comprehensive and accurate overview of user interests from these representations, this embodiment employs average pooling. Average pooling is an operation that averages multiple features, effectively fusing interest information from different parts to obtain a comprehensive user representation. Finally, this application uses a small MLP to obtain the user representation u. For a single-objective, multi-behavior recommender problem, maximizing Q(u) is the optimization objective of this application.

[0143] Where v t ∈v L+1 This represents the item whose interaction with the target behavior is to be predicted.

[0144] See Figure 4 This application also provides a multi-behavior sequence recommendation system based on graph neural networks, including:

[0145] User multi-behavior sequence acquisition unit, used to acquire user multi-behavior sequences;

[0146] The user interaction graph acquisition unit is used to input multiple user behavior sequences into a pre-constructed global item graph to acquire the user interaction graph.

[0147] A multi-behavior sequence recommendation analysis unit is used to analyze and mine user interaction graphs using a pre-built graph-perceptron hybrid model to obtain recommendation results. The graph-perceptron hybrid model includes:

[0148] The interest extraction module divides the user interaction graph into overlapping interest subgraphs.

[0149] The interest embedding module transforms the interest subgraph into an initial interest representation, which is then input into the interest mining module.

[0150] The interest mining module generates the final interest representation by modeling each interest information and the relationship between interests through multiple residual networks, based on the initial interest representation.

[0151] The user embedding module obtains the user representation by global average pooling based on the final interest representation, and then connects it to the fully connected layer for the next item prediction.

[0152] Specific limitations regarding the multi-behavior sequence recommendation system based on graph neural networks can be found in the limitations of the multi-behavior sequence recommendation method based on graph neural networks mentioned above, and will not be repeated here. Each unit in the aforementioned multi-behavior sequence recommendation system based on graph neural networks can be implemented entirely or partially through software, hardware, or a combination thereof. These units can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each unit.

[0153] The following describes the experimental process and results of the performance and effectiveness of the graph-perceptron hybrid model disclosed in this application.

[0154] To accurately measure and demonstrate the performance of the recommendation system proposed in this application, we conducted in-depth experiments on three industry-recognized large-scale real-world datasets. These datasets are highly authoritative in the field of multi-behavioral sequence recommendation and have been widely adopted as standard benchmarks for performance evaluation. They cover diverse behavioral patterns generated by users when interacting with online platforms, including but not limited to browsing, favorites, adding to cart, and final purchase behavior. Specifically, the IJCAI dataset was released by the IJCAI competition 2015 and is used for the task of predicting repeat buyers. The CIKM and Taobao datasets were released by Alibaba and Taobao, Chinese e-commerce companies, which are among the largest e-commerce platforms in China, with hundreds of millions of users and massive amounts of transaction data. By conducting experiments on such representative datasets, we can ensure that the recommendation algorithm proposed in this application is not only theoretically innovative but also effective in practical applications, meeting user needs and improving their shopping experience.

[0155] To fully validate the effectiveness of the GraphMLP-Mixer method, this paper conducts a series of detailed experiments and compares it in depth with a range of state-of-the-art benchmark methods. These benchmark methods cover multiple domains, from traditional sequence recommendation to multi-benchmark sequence recommendation, to ensure that the comparisons in this paper are comprehensive and thorough.

[0156] In exploring the performance evaluation of recommender systems, this paper selects two widely accepted evaluation metrics: Hit Ratio (HR@K) and Normalized Discounted Cumulative Gain (NDCG@K).

[0157] Table 1 presents a comprehensive performance comparison across different datasets and summarizes the following observations: First, multi-behavior methods generally outperform general sequence recommendation methods, highlighting the significance of the fundamental research question addressed in this application. By using multiple behavioral data as auxiliary information to model the complex relationships between different user behaviors, this application effectively transforms general preferences into specific behaviors to improve performance. Second, comparing the results of GNN-based methods with traditional methods, it can be concluded that graph network-based methods generally perform better than methods based on simple attention.

[0158]

[0159]

[0160] This further demonstrates the powerful modeling capabilities of graph neural networks, which can represent higher-order information about users and items. Finally, GraphMLP-Mixer consistently outperforms other baselines across all datasets, showcasing its effectiveness and generalization ability. GraphMLP-Mixer effectively captures the global dependencies of user behavior and mitigates the problem of information over-compression. By integrating information from the global item graph and leveraging the multi-dimensional information mining capabilities of MLP-Mixer, this model achieves a deep understanding and representation of user behavior sequences. This design gives GraphMLP-Mixer a significant advantage in handling multi-behavior sequence recommendation problems.

[0161] Table 1 compares the overall performance of all methods in terms of HR@K and NDCG@K (K = 10, 20). (p < 0.05)

[0162] This application presents a novel and efficient graph neural network architecture—GraphMLP-Mixer—with two significant advantages: first, it can effectively capture the global dependencies of user behavior while mitigating the problem of information over-compression; second, its time and space efficiency are significantly improved, and its complexity is linearly related to the number of user interaction behaviors, outperforming existing GNN-based multi-behavior sequence recommendation models. Experimental results show that GraphMLP-Mixer outperforms existing GNN-based multi-behavior recommendation methods on three real-world public datasets, achieving significant performance improvements while effectively reducing model complexity, demonstrating its great potential in practical applications.

[0163] This invention is not limited to the specific embodiments described above. Any modifications made by those skilled in the art based on the above concept without creative effort are within the scope of protection of this invention.

Claims

1. A multi-behavior sequence recommendation method based on graph neural networks, characterized in that, Includes the following steps: Obtain user multi-behavior sequences; Input the user's multi-behavior sequence into a pre-constructed global item graph to obtain the user interaction graph; By using a pre-built graph-perceptron hybrid model, user interaction graphs are analyzed and mined to obtain recommendation results; The pre-constructed global item graph: Nodes in the graph represent the set of item nodes appearing in the interaction sequence; edges in the graph connect any two item nodes. and Weight of the edges between It is the sum of the number of times the two nodes are connected in all sequences at order k = [1,2,...,n]. Indicates the spacing between adjacent items; ; ; Represents two item nodes and The weight of the edges between them. Indicates to Normalized weights; Represents items Position in the sequence, This indicates that the difference between two items in the sequence is... Number of times, It is a collection of users. User The interaction sequence; The data processing steps of the graph-perceptron hybrid model include: Interest extraction steps: Divide the user interaction graph into overlapping interest subgraphs; Interest embedding steps: Transform the interest subgraph into an initial interest representation; Interest mining steps: For the initial interest representation, the final interest representation is generated by modeling each interest information and the relationship between interests through multiple residual networks; User embedding steps: For the final interest representation, a user representation is obtained through global average pooling and connected to a fully connected layer for the next item prediction.

2. The multi-behavior sequence recommendation method based on graph neural networks according to claim 1, characterized in that, The construction process of the global item graph also includes: for edges with weights greater than a preset threshold, their weights are reset to 0.

3. The multi-behavior sequence recommendation method based on graph neural networks according to claim 1, characterized in that, The interest extraction step is implemented using an improved method based on METIS. This improved METIS method, i.e., the interest extraction step, includes: Using METIS, the user interaction graph G is divided into P non-overlapping tiles. ,and , , ; Extend the tiles to their one-hop neighbors to preserve information about the links between tiles and utilize all graph edges.

4. The multi-behavior sequence recommendation method based on graph neural networks according to claim 1, characterized in that, The interest embedding step involves converting the interest subgraph into an initial interest representation, including: (51) For each interest subgraph, for the original nodes ∈ and edge ∈ Perform linear embedding and linear projection onto d-dimensional hidden features: ； ； in, , , , , , , , These are learnable parameters; (52) Graph update is performed using graph convolutional layers with MP-GNN: Nodes and edges are represented by each tile. Apply MP-GNN for updates; ； ； in, , , , , , It is a layer index. It is a tile index. , Represents a node. It is a node The neighborhood; a function with learnable parameters. , ; , , which are the average representations of the tile nodes and tile edges, respectively; , It is an MLP-based function that operates on and ; For each node and edge that covers multiple tiles due to tile overlap, update the node / edge representation by averaging the representations in the overlapping tiles: ； ； in , These are the coverage nodes. ,side The set of all tiles; (53) All node vectors in the array are average pooled, so that... And apply a small MLP to obtain fixed-size tile embeddings. .

5. The multi-behavior sequence recommendation method based on graph neural networks according to claim 4, characterized in that, The interest embedding step, which converts the interest subgraph into an initial interest representation, further includes: adding location encoding information to nodes and tiles, wherein the location encoding includes: (61) Node location code PE Introduction ∈ and a learnable matrix ∈ Enhance the information of the input node features. ； (62) Plot location code PE Add tile location encoding to the tile embedding of different tiles Obtain new tile embeddings, including: (621) Through the coarsened adjacency matrix of the interest subgraph ∈ To capture relative position information: ,in, It is a standard graph cutting operator that computes interest subgraphs. , The number of connecting edges between them; (622) Based on coarsened adjacency matrix Get the first tile Location encoding of each tile ∈ Patch-based location encoding Update the embedded representation of the tile ,in, It is an embedded representation of a tile.

6. The multi-behavior sequence recommendation method based on graph neural networks according to claim 1, characterized in that, The interest mining steps involve: for the initial interest representation, generating the final interest representation by modeling each interest information and the relationships between interests using multiple residual networks, including: processing the interest embedding set using MLP-Mixer. Embedding collections for interests , The steps of interest discovery are represented as follows: ； ； Among them, symbols It is represented as the GELU nonlinear activation function. Representation layer normalization operation, matrix , , , ,in, and These are the adjustable hidden layer widths in token-mixing and channel-mixing MLPs, respectively.

7. The multi-behavior sequence recommendation method based on graph neural networks according to claim 1, characterized in that, The user embedding step involves: obtaining a user representation through global average pooling based on the final interest representation, and connecting it to a fully connected layer for the next item prediction, including: Average pooling steps: ;in, , It is a binary variable, with a value of 1 for non-empty blocks and a value of 0 for empty blocks; Prediction results are obtained based on a fully connected layer MLP. .

8. The multi-behavior sequence recommendation method based on graph neural networks according to claim 1, characterized in that, The interest extraction step further includes: For the original image Create a perturbation graph ,in By randomly deleting the original edge set A small portion of the edges are obtained; exist An improved METIS algorithm is applied to obtain different item node partitions. .

9. A multi-behavior sequence recommendation system based on graph neural networks, characterized in that, include: User multi-behavior sequence acquisition unit, used to acquire user multi-behavior sequences; The user interaction graph acquisition unit is used to input a user multi-behavior sequence into a pre-constructed global item graph to acquire the user interaction graph; the pre-constructed global item graph: the nodes in the graph represent the set of item nodes appearing in the interaction sequence; any two item nodes on the edge of the graph... and Weight of the edges between It is the sum of the number of times the two nodes are connected in all sequences at order k = [1,2,...,n]. Indicates the spacing between adjacent items; ; ; Represents two item nodes and The weight of the edges between them. Indicates to Normalized weights; Represents items Position in the sequence, This indicates that the difference between two items in the sequence is... Number of times, It is a collection of users. User The interaction sequence; A multi-behavior sequence recommendation analysis unit is used to analyze and mine user interaction graphs using a pre-built graph-perceptron hybrid model to obtain recommendation results. The graph-perceptron hybrid model includes: The interest extraction module divides the user interaction graph into overlapping interest subgraphs. The interest embedding module transforms the interest subgraph into an initial interest representation, which is then input into the interest mining module. The interest mining module generates the final interest representation by modeling each interest information and the relationship between interests through multiple residual networks, based on the initial interest representation. The user embedding module obtains the user representation by global average pooling based on the final interest representation, and then connects it to the fully connected layer for the next item prediction.

Images