Adaptive anchor point diagram attention network embedding method and system for large-scale sparse bipartite graph

Abstract

The invention discloses a large-scale sparse bipartite graph-oriented adaptive anchor point diagram attention network embedding method and system, and the method comprises the steps: firstly, building a dynamic anchor point number generation mechanism through analyzing the scale, sparseness, user node number, average degree and other graph characteristics of a graph, and achieving the adaptive adjustment of the number of anchor points; secondly, a full-connection attention mechanism is designed, efficient information transmission between a user node and an anchor node is achieved through bidirectional attention operation of FCAsend and FCArecv, and the problem of information loss of a sparse region is solved; and finally, a bipartite graph attention network with entropy regularization is introduced, a uniform weight matrix and a sparse attention mechanism are adopted to process a dense region, and node features are effectively prevented from being excessively smooth. According to the method, through innovative combination of the self-adaptive anchor points and the double attention mechanism, the embedding quality of the large-scale sparse bipartite graph is remarkably improved, and the performance superior to that of an existing method is achieved in actual tasks such as a recommendation system and link prediction.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology, and in particular to an adaptive anchor graph attention network embedding method and system for large-scale sparse bipartite graphs. Background Art

[0002] Bipartite graphs are widely used data structures for modeling the relationship between two different types of entities. They are widely used in fields such as recommender systems, social network analysis, bioinformatics, and knowledge graph queries. Bipartite graph embedding aims to map each node into a compact low-dimensional vector, which is crucial for the performance of downstream applications. Currently, research on bipartite graph embedding focuses on methods such as matrix decomposition, random walks, and graph neural networks. Matrix decomposition methods reveal latent features by decomposing the original matrix into two low-rank matrices, but in sparse graphs, the measurement is unreliable due to insufficient node interactions. Random walk methods capture structural information by simulating random walk paths between nodes, but the embedding effect is greatly affected by the coverage of path generation. Graph neural network methods use multi-layer neural network architectures to learn complex node features and relationships, but the computational overhead is large when processing large-scale sparse bipartite graphs.

[0003] In recent years, anchor-based methods have gradually attracted attention. This type of method uses anchor-assisted information transfer to improve the embedding quality, but the embedding quality is greatly affected by the number of anchors. However, current bipartite graph embedding methods still face severe challenges when dealing with large-scale sparse graphs. Statistics show that when the number of edges reaches millions, the edge density is generally lower than 10^-5. This extreme sparsity leads to a lack of sufficient contextual information for nodes. In sparse areas, due to limited node connections, existing methods find it difficult to accurately capture the functions and roles of nodes, affecting the performance of downstream tasks. In dense areas, multi-layer information propagation leads to the convergence of node features, reducing the distinctiveness of the embedding representation and affecting the model's ability to identify different nodes.

[0004] In addition, existing methods also have obvious deficiencies in scalability. As the size of the graph increases, the computational overhead increases dramatically, making it difficult to maintain high computational efficiency while ensuring embedding quality, which severely limits its deployment in practical applications. Therefore, there is an urgent need for a bipartite graph embedding method that can simultaneously solve the problems of missing context in sparse regions and over-smoothing in dense regions and has good scalability. Summary of the invention

[0005] The present invention proposes solutions to the main problems existing in existing bipartite graph embedding methods when processing large-scale sparse graphs.

[0006] Specifically:

[0007] Solve the problem of missing context information in sparse areas. Existing methods are difficult to effectively capture the global context information of nodes in sparse areas, resulting in insufficient node representation and affecting the performance of downstream tasks. This paper designs an adaptive anchor point generation mechanism and a fully connected attention mechanism to enhance information transfer between nodes and effectively obtain global context information.

[0008] Alleviate the problem of over-smoothing in dense areas. Existing methods easily lead to convergence of node features in the multi-layer propagation process, reducing the distinctiveness of representation. The present invention proposes a bipartite graph attention network based on entropy regularization, which maintains the uniqueness of node representation by limiting the excessive concentration of attention weights.

[0009] Achieve a balance between computational efficiency and embedding quality. Existing methods have high computational overhead when processing large-scale graphs, and it is difficult to strike a balance between efficiency and quality. The present invention improves computational efficiency while ensuring embedding quality through optimized network architecture design and efficient message passing mechanism.

[0010] The technical solution adopted by the present invention is as follows:

[0011] The system of the present invention mainly includes three core modules:

[0012] Adaptive anchor generation module: This module adaptively generates an appropriate amount of anchor nodes by analyzing the scale, sparsity, number of nodes, average degree and other characteristics of the graph. These anchor nodes act as information transfer stations, significantly improving the efficiency of information transmission in large-scale sparse graphs.

[0013] Message Passing Network Module: This module implements efficient information transmission based on the fully connected attention mechanism. Through carefully designed steps such as dimension conversion, anchor projection, user reception and information aggregation, it ensures that information can be effectively transmitted between different types of nodes.

[0014] Bipartite graph attention network module: This module is dedicated to processing the interaction between users and items. Through unified feature transformation and mask attention mechanism, it not only maintains computational efficiency but also ensures the quality of node representation. At the same time, the introduced entropy regularization mechanism effectively prevents feature degradation.

[0015] The method of the present invention is as follows:

[0016] Step 1: Input data preprocessing and feature analysis This step is the basic part of the entire method. Through systematic data preprocessing and multi-dimensional feature extraction, it lays a solid foundation for subsequent model construction.

[0017] Step 101: Data acquisition and preprocessing. This step first normalizes the original bipartite graph data and constructs it into a standard mathematical representation G = (U, V, E). U represents the user node set, which includes all user entities in the system; V represents the item node set, which covers all interactive item entities; and E is the edge set, which records all historical interaction information between users and items. This mathematical representation not only facilitates subsequent calculations and processing, but also ensures the structure and integrity of the data.

[0018] Step 102: Multi-dimensional feature analysis This step comprehensively analyzes the graph structure through four key dimensions to fully understand the characteristics of the graph;

[0019] 1021: Graph scale calculation - using logarithmic form S = log (n V +n U ) The scale index of the computation graph, where n V and n U Represent the number of items and users respectively; this logarithmic design can effectively handle graph structures of different magnitudes, making the calculation results more reasonable and comparable.

[0020] 1022: Sparsity evaluation - through the formula C = |E| / (n U ×n V ) calculates the sparsity of the graph, where |E| represents the number of actual edges. This indicator intuitively reflects the connection density of the graph and has important guiding significance for subsequent algorithm design.

[0021] 1023: User scale statistics - record the total number of users in the system N = n U This basic but critical indicator directly affects the subsequent computational complexity and system performance.

[0022] 1024: Activity Analysis - Calculate the average user connection degree D = Σdegree(u i ) / n U , by analyzing the average number of interactions for each user, we can gain an in-depth understanding of the user activity distribution.

[0023] Step 103: Calculation of the number of adaptive anchor points Based on the above feature analysis results, an innovative formula for calculating the number of adaptive anchor points is designed: This formula cleverly combines the four key features of the graph: scale (S), sparsity (C), number of users (N), and average degree (D), and balances the influence of each feature through adjustable weight parameters α, β, and γ. The use of logarithmic form and power of 2 ensures the discreteness of the calculation results and ensures that the number of anchor points can be dynamically adjusted with the characteristics of the graph. This adaptive design greatly improves the model's adaptability to graph structures of different scales and characteristics.

[0024] Step 2: Construction of adaptive anchor point generation module This step focuses on the construction and initialization of anchor nodes, which is a key link in achieving efficient information transfer.

[0025] Step 201: Anchor embedding initialization adopts a carefully designed three-stage initialization strategy:

[0026] Step 2011: Use Xavier normal distribution in Generate initial values ​​within the range. This initialization method can effectively avoid the gradient vanishing and gradient exploding problems, making the training process more stable.

[0027] Step 2012: Construct anchor embedding matrix A∈R^(n A ×d), where each row represents a d-dimensional vector representation of an anchor point. This matrix, as an intermediary for information transmission, plays a vital role in subsequent calculations.

[0028] Step 2013: Normalize the generated embedding matrix to ensure the stability of the numerical distribution through the standardization operation, providing a good numerical basis for subsequent attention calculations.

[0029] Step 202: Node embedding initialization adopts a dual embedding representation strategy:

[0030] Step 221: Initialize the user node embedding matrix U∈R^(n U ×d), constructing a unique feature representation for each user.

[0031] Step 2022: Initialize the item node embedding matrix V∈R^(n V ×d), capturing the latent characteristics of each item.

[0032] Step 2023: Perform normalization operations on all initialized matrices to ensure that the features of different types of nodes are distributed within the same numerical range to facilitate subsequent feature fusion and calculation.

[0033] Step 3: Design and implementation of fully connected attention mechanism In this step, efficient information transmission between nodes is achieved through innovative attention mechanism design.

[0034] Step 301: The attention score calculation adopts a multi-level attention calculation strategy:

[0035] Step 3011: Construct a feature transformation matrix W∈R^(d′×d) and an attention vector a∈R^(2d′) for feature projection and attention calculation.

[0036] Step 3012: Through the composite function e ij =LeakyReLU(aT [Wu i ||Wa j ]) Calculate the attention scores between node pairs, where the use of LeakyReLU can effectively avoid the gradient vanishing problem.

[0037] Step 3013: Use the softmax function to normalize the attention score.

[0038]

[0039] Get the final attention weights, ensuring that the sum of the weights is 1 and non-negative.

[0040] Step 302: Message passing implementation designs a four-stage message passing mechanism:

[0041] Step 3021: Dimension conversion - using fully connected attention to generate initial message S = FCA send (U) realizes the mapping transformation of feature space.

[0042] Step 3022: Anchor point projection - through matrix operation R = SA T Project the messages into the anchor point space to achieve effective aggregation of information.

[0043] Step 3023: User receives, processes and converts the received message A' = FCA recv (R) to ensure effective delivery of information.

[0044] Step 3024: Information aggregation - information fusion is achieved through residual connection and nonlinear activation H'=H+sin(A'), where the introduction of the sin function increases the nonlinear expression ability of the model.

[0045] Step 4: Bipartite Graph Attention Network Construction This is a key step specifically designed to handle user-item interactions.

[0046] Step 401: Feature conversion adopts a unified feature conversion strategy:

[0047] Step 4011: Construct a unified weight matrix W to ensure the consistency and interpretability of feature transformation.

[0048] Step 4012: Convert user node features Maintain the expressiveness of user characteristics.

[0049] Step 4013: Convert item node features Ensure effective representation of item characteristics.

[0050] Step 402: Masked attention mechanism implements efficient sparse attention calculation:

[0051] Step 4021: Construct an adjacency matrix to accurately represent the actual connection relationship between nodes.

[0052] Step 4022: Dynamically set the attention mask based on the actual existence of the edge: keep the original attention score e for the existing edge ij , and the non-existent edges are set to negative infinity -∞, thus achieving efficient sparse computing.

[0053] Step 403: Multiple loss functions are designed to achieve comprehensive optimization goals:

[0054] Step 4031: Prediction loss construction - using cross entropy loss L fit Evaluate model predictive ability.

[0055] Step 4032: Regularization loss introduction - L2 regularization reg Prevent overfitting.

[0056] Step 4033: Entropy regularization application - introducing entropy regularization loss L a Maintain a diverse distribution of attention.

[0057] Step 4034: Total loss function combination - through L total =L fit +L reg +L a Achieve multi-objective optimization.

[0058] Step 5: Model training optimization This step improves model performance through multiple optimization strategies.

[0059] Step 501: Batch normalization is implemented using a three-stage normalization strategy:

[0060] Step 5011: Calculate batch statistics, including mean and variance.

[0061] Step 5012: Perform standardization processing to ensure stable data distribution.

[0062] Step 5013: Introduce learnable parameters to enhance the adaptability of the model.

[0063] Step 502: Residual connection construction to achieve efficient information transfer:

[0064] Step 5021: retain the original feature information.

[0065] Step 5022: Normalize the converted features.

[0066] Step 5023: Connect h″ via residual i =BatchNorm(h′ i )+hi Fusion information.

[0067] Step 503: Parameter optimization design adopts multiple optimization strategies:

[0068] Step 5031: Use the Adam optimizer to update parameters.

[0069] Step 5032: Implement dynamic adjustment of learning rate.

[0070] Step 5033: Introduce an early stopping mechanism to avoid overfitting.

[0071] Step 6: Node Embedding Representation Generation This step is responsible for generating the final node representation.

[0072] Step 601: User node embedding generation

[0073] Step 6011: Integrate the global information from the anchor points to ensure the globality of the representation.

[0074] Step 6012: Fuse local interaction information to capture the user's personalized features.

[0075] Step 6013: Generate the final user embedding U'∈R^(n U ×d), achieving high-quality user representation.

[0076] Step 602: Item node embedding generation

[0077] Step 6021: Integrate the feature information of the bipartite graph attention network.

[0078] Step 6022: Enhance representation capability through nonlinear transformation.

[0079] Step 6023: Generate item node embedding V'∈R^(n V ×d) to achieve effective representation of items.

[0080] Step 7: Downstream Task Application This step applies the generated embeddings to actual tasks.

[0081] Step 701: Recommended task implementation

[0082] Step 7011: Calculate the user-item similarity score based on the embedding, using a metric such as cosine similarity.

[0083] Step 7012: Sort the similarity scores of all candidate items in descending order.

[0084] Step 7013: Select the K items with the highest scores as recommendation results.

[0085] Step 702: Link prediction implementation

[0086] Step 7021: Calculate similarity index of node pairs using node embedding.

[0087] Step 7022: Set an appropriate threshold for connection judgment.

[0088] Step 7023: Output the final prediction result, including the probability of the connection existing and the binary prediction label.

[0089] The beneficial effects created by the present invention are: effectively solving the problem of information transmission in sparse areas: innovatively proposing an adaptive anchor point generation mechanism, dynamically adjusting the number of anchor points according to the characteristics of the graph, and enhancing the information exchange capability between nodes through the fully connected attention mechanism. At the same time, the introduction of entropy regularization effectively prevents the phenomenon of feature homogeneity. A balance between computational efficiency and scalability is achieved: the use of a masked attention mechanism reduces computational complexity, the number of model parameters is reduced by a unified weight matrix, and the use of batch normalization and residual connections accelerates the model convergence process. The robustness of the model is guaranteed: the design of a multi-objective loss function improves the generalization ability of the model, the adaptive mechanism enables the model to better handle graphs with different features, and the use of L2 regularization effectively prevents overfitting. It has a wide range of practical application value: the present invention is applicable to multiple fields such as recommendation systems and social network analysis, and can effectively process sparse bipartite graphs with a scale of millions of nodes, with good scalability and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 is a flow chart of the method of the present invention;

[0091] Figure 2 It is a structural diagram of the model of the present invention. DETAILED DESCRIPTION

[0092] The technical solutions in the implementation cases of the present invention are described in detail below in conjunction with the accompanying drawings. The following implementations are applicable to the present invention, but are not intended to limit the scope of the present invention.

[0093] like Figure 1 and Figure 2 As shown in the figure, the adaptive anchor graph attention network proposed in this invention adopts a modular design concept and consists of three closely cooperating core functional modules: adaptive anchor generation module, message passing network module and bipartite graph attention network module. This modular architecture design not only improves the maintainability and scalability of the system, but more importantly, it can be specialized for different characteristics and challenges of large-scale sparse bipartite graphs.

[0094] Each module is carefully designed to solve specific technical difficulties: the adaptive anchor generation module focuses on solving the problem of information transmission efficiency; the message passing network module focuses on achieving effective information exchange between nodes; and the bipartite graph attention network module is committed to optimizing the modeling of user-item interaction relationships. These three modules work together through well-designed interfaces and data flows to build a complete solution.

[0095] Step 1: The input bipartite graph is denoted as G = (U, V, E), and its specific composition is:

[0096] User node set U={u1,u2,...u i}: Contains all user entities in the system, and each user node has a unique identifier.

[0097] Item node set V = {v1,v2,...,v j}: Covers all interactive item entities and also ensures that each item has a unique identifier.

[0098] Interaction Edge Set It records all historical interaction information between users and items, which may include additional information such as interaction type and timestamp.

[0099] Step 101: Analyze four key features of the graph structure:

[0100] Step 1011: Scale index S - using S = log (n U +n V ) to calculate the size of the graph, where n U and n V Represent the number of users and the number of items respectively. This logarithmic form can effectively handle graph structures of different magnitudes and produce more reasonable and comparable results.

[0101] Step 1012: Sparsity C - by formula C = |E| / (n U *n V ) calculates the sparsity of the graph, where |E| is the actual number of edges. This indicator intuitively reflects the connection density of the graph and has important guiding significance for subsequent algorithm design.

[0102] Step 1013: Number of users N-total number of users in the direct recording system N=n U ,This basic but critical feature directly affects the computation complexity and ,system performance.

[0103] Step 1014: Average degree D-calculate the average user connection degree D = Σdegree(u) / n U , and understand the distribution of user activity levels.

[0104] Step 1015: Calculate the number of adaptive anchor points according to the formula: This formula cleverly combines the scale, sparsity, number of users and average degree of the graph, and balances the influence of each feature through adjustable weight parameters α, β, and γ. The use of logarithmic form and the power of 2 ensures the discreteness of the calculation results and ensures that the number of anchor points can be dynamically adjusted according to the characteristics of the graph.

[0105] Step 2: Adaptive anchor point generation With the above in-depth analysis of the graph structure, we will proceed to build an adaptive anchor point module. The main task of this module is to generate an appropriate number of anchor nodes and initialize them.

[0106] Step 201: Use Xavier normal distribution to Initialize the anchor embedding matrix A∈R^(n A ×d). This initialization method can effectively avoid the gradient disappearance and explosion problems, making the subsequent training process more stable.

[0107] Step 202: Normalize the generated embedding matrix to ensure the stability of the numerical distribution and lay a good foundation for subsequent attention calculations.

[0108] Step 3: Initialize user node embedding U∈R^(n U ×d) and item node embedding V∈R^(n V ×d). The former constructs a unique feature representation for each user, while the latter captures the latent features of each item.

[0109] Step 303: Normalize the two embedding matrices to ensure that the features of nodes of different types are distributed in the same numerical range, which is beneficial to subsequent feature fusion and calculation.

[0110] Step 4: Fully connected attention mechanism With the initialized node and anchor embedding, design a four-stage message passing mechanism

[0111] Step 401: Construct the feature transformation matrix W∈R^(d'×d) and the attention vector a∈R^(2d') to project the node features into a higher-dimensional space and calculate the attention scores between nodes. Specifically,

[0112] Step 402: Use formula e ij =LeakyReLU(a T [Wu i ||Wa j ]) Calculate the attention score e between user node i and anchor node j ij j. The introduction of the LeakyReLU activation function here can effectively avoid the gradient vanishing problem.

[0113] Step 403: Use the softmax function to normalize the attention score to obtain the final attention weight This ensures that the weight sum is 1 and non-negative, laying the foundation for subsequent feature aggregation.

[0114] Step 404: Dimension conversion - generating initial message S = FCA using fully connected attention send (U) realizes the mapping transformation of feature space.

[0115] Step 405: Anchor point projection - via R = SA T Project the messages into the anchor point space to achieve effective aggregation of information.

[0116] Step 406: User receives, processes and converts the received message A' = FCA recv (R), ensuring that information can be effectively delivered back to the user node.

[0117] Step 407: Information aggregation - information is fused through residual connection and sin activation H'=H+sin(A') to enhance the nonlinear expression ability of the model.

[0118] Step 5: Further process the interaction between users and items

[0119] Step 501: Use a unified feature conversion matrix W to ensure the consistency and interpretability of user and item feature conversion. Specifically, the user features are converted into Item features are converted to Ensure that the two types of node features are processed in the same unified space.

[0120] Step 502: Dynamically set the attention mask according to the existence of edges: For node pairs with edges, retain the original attention score e ij ; For node pairs without edges, the attention score is set to negative infinity -∞.

[0121] Step 503: In terms of loss function design, three types of losses are integrated: prediction loss L fit , L2 regularization loss L reg And entropy regularization loss La. Among them, L fit Cross entropy calculation was used to evaluate the predictive ability of the model; L reg L2 regularization is used to prevent overfitting; the entropy regularization loss La introduced helps maintain the diversity of attention distribution and avoid over-concentration. These three losses are combined into a total loss L total =L fit +L reg +La, can realize multi-objective optimization of the model.

[0122] Step 6: Model training optimization During the model training optimization process, a series of techniques are used to improve performance and stability.

[0123] Step 601: Batch normalization mechanism is applied, including: calculating batch statistics (mean and variance), standardization to ensure data distribution stability, and introducing learnable parameters to enhance model adaptability.

[0124] Step 602: A residual connection mechanism is constructed. This includes: retaining the original feature information, normalizing the transformed features, and finally i =BatchNorm(h′ i )+h i Fusion information. This can not only accelerate the convergence of the model, but also effectively alleviate the gradient disappearance problem.

[0125] Step 603: The Adam algorithm is used, and the dynamic adjustment of learning rate and early stopping mechanism are implemented. These strategies can not only improve the model performance, but also effectively prevent overfitting.

[0126] Step 7: Node embedding representation Based on the aforementioned network architecture and training optimization, high-quality user and item embedding representations are finally generated.

[0127] Step 701: For user embedding U'∈R^(n U ×d), which combines the global information and local interaction information from the anchor points to ensure the global and personalized characteristics of the representation.

[0128] Step 702: For the item embedding V'∈R^(n V ×d), which makes full use of the feature information learned by the bipartite graph attention network and further enhances the representation ability through nonlinear transformation.

[0129] Step 8: Downstream Task Application With the high-quality node embeddings generated above, they can be applied to various practical tasks, including recommendation systems and link prediction.

[0130] Step 801: In the recommendation task, the scores are calculated based on the similarity of user-item embeddings, all candidate items are sorted in descending order, and finally the top K items with the highest scores are selected as the recommendation results.

[0131] Step 802: In the link prediction task, the node embedding is used to calculate the similarity index of the node pair, and the presence of a connection is determined according to the set threshold, and the final prediction result is output, including the connection probability and the binary label.

Claims

1. An adaptive anchor graph attention network embedding method for large-scale sparse bipartite graphs, characterized in that: The steps are: Step 1: Input data preprocessing and feature analysis: Step 101: Data acquisition and preprocessing: First, the original bipartite graph data is normalized and constructed into a standard mathematical representation G = (U, V, E); where U represents the user node set, V represents the item node set, and E represents the edge set; Step 102: Multi-dimensional feature analysis: This step obtains the analysis results of the graph structure and graph features through graph scale calculation, sparsity evaluation, user scale statistics, and activity analysis; Step 103: Calculate the number of adaptive anchor points: Based on the analysis results, calculate the number of adaptive anchor points: Among them, S is the scale of the graph, C is the sparsity, N is the number of users, and D is the average degree. By adjusting the weight parameters α, β, and γ, the influence of each feature is balanced; Step 2: Construct an adaptive anchor point generation module and perform anchor point embedding and node embedding initialization; Step 3: Construct a fully connected attention mechanism, calculate the attention score between the user node and the anchor node through the fully connected attention mechanism, and determine the four-stage message passing mechanism; Step 4: Build a bipartite graph attention network: Step 401: Adopt a unified feature conversion strategy, construct a unified weight matrix W, and convert the user node features Maintain the expressiveness of user features and transform item node features Ensure effective representation of item characteristics; Step 402: Construct an adjacency matrix and dynamically set the attention mask based on the actual existence of edges: keep the original attention score e for the existing edges ij , for non-existent edges, set to negative infinity -∞; Step 403: Design multiple loss functions L total =L fit +L reg +L a To achieve multi-objective optimization, L fit is the cross entropy loss, L reg Apply L2 regularization, L a is the entropy regularization loss; Step 5: Train and optimize the model; Step 6: Generate node embedding representation: including user node embedding generation and item node embedding generation; Step 7: Downstream task application: Calculate the user-item similarity score based on embedding, use cosine similarity and other metrics to sort the similarity scores of all candidate items in descending order, and select the K items with the highest scores as the recommendation results; use node embedding to calculate the similarity index of node pairs, set appropriate thresholds for connection judgment, and output the final prediction results, including the probability of connection existence and binary prediction labels.

2. The method of embedding an adaptive anchor graph attention network for large-scale sparse bipartite graphs according to claim 1, characterized in that: In the step 102, the specific method is: A comprehensive analysis of the graph structure is conducted through four key dimensions: 1021: Graph scale calculation: using logarithmic form S = log(n V +n U ) The scale index of the computation graph, where n V and n U Represent the number of items and users respectively, and process graph structures of different magnitudes in logarithmic form; 1022: Sparsity evaluation: by formula C = |E| / (n U ×n V ) Calculate the sparsity of the graph, where |E| represents the actual number of edges and the sparsity C reflects the connection density of the graph; 1023: User scale statistics: record the total number of users in the system N = n U ,The total number of users N affects the subsequent computational complexity and system performance; 1024: Activity analysis: Calculate the average user connection degree D = Σdegree(u i ) / n U ,By analyzing the average number of interactions of each user, we can obtain the distribution of user activity.

3. The method of embedding an adaptive anchor graph attention network for large-scale sparse bipartite graphs according to claim 1, characterized in that: In the step 2, the specific method is: Step 201: Anchor embedding initialization, using a carefully designed three-stage initialization strategy: Using Xavier normal distribution Generate initial values ​​within the range and construct the anchor embedding matrix A∈R^(n A ×d), where each row represents a d-dimensional vector representation of an anchor point. This matrix acts as an intermediary for information transmission and normalizes the generated embedding matrix; Step 202: Initialize node embedding, using a dual embedding representation strategy: Initialize the user node embedding matrix U∈R^(n U ×d), construct feature representation for each user, and initialize the item node embedding matrix V∈R^(n V ×d), capturing the potential features of each item and performing normalization operations on all initialized matrices to ensure that the features of different types of nodes are distributed in the same numerical range.

4. The method of embedding an adaptive anchor graph attention network for large-scale sparse bipartite graphs according to claim 1, characterized in that: In the step 3, the specific method is: Step 301: Attention score calculation, using a multi-level attention calculation strategy: constructing a feature transformation matrix W∈R^(d′×d) and an attention vector a∈R^(2d′) for feature projection and attention calculation, and using the composite function e ij =LeakyReLU(a T [Wu i ||Wa j ]) Calculate the attention scores between node pairs to avoid the gradient vanishing problem, and use the softmax function to normalize the attention scores. Get the final attention weights, ensuring that the sum of the weights is 1 and non-negative; Step 302: Implementation of message delivery, design of a four-stage message delivery mechanism: Dimension conversion: Generate initial message S=FCA using fully connected attention send (U), realize the mapping transformation of feature space; Anchor point projection: through matrix operation R = SA T Project the message into the anchor point space to achieve effective aggregation of information; User reception: Process and convert the received message A' = FCA recv (R) ensure effective delivery of information; Information aggregation: Information fusion is achieved through residual connection and nonlinear activation H'=H+sin(A'), where the introduction of the sin function increases the nonlinear expression ability of the model.

5. The method of embedding an adaptive anchor graph attention network for large-scale sparse bipartite graphs according to claim 1, characterized in that: In the step 5, the specific method is: Step 501: Batch normalization is implemented using a three-stage normalization strategy: calculating batch statistics, including mean and variance; performing standardization to ensure stable data distribution; Introducing learnable parameters to enhance the adaptability of the model; Step 502: Residual connection construction to achieve efficient information transfer: retain the original feature information, normalize the transformed features, and connect h″ through the residual connection i =BatchNorm(h′ i )+h i Fusion of information; Step 503: Parameter optimization design, using multiple optimization strategies: using the Adam optimizer to update parameters, implementing dynamic adjustment of the learning rate, and introducing an early stopping mechanism to avoid overfitting.

6. The method of embedding an adaptive anchor graph attention network for large-scale sparse bipartite graphs according to claim 1, characterized in that: In the step 6, the specific method is: Step 601: User node embedding generation: integrating global information from anchor points to ensure the globality of representation; fusing local interaction information to capture user personalized features; Generate the final user embedding U'∈R^(n U ×d), achieving high-quality user representation; Step 602: Generate item node embedding: Integrate the feature information of the bipartite graph attention network, enhance the representation capability through nonlinear transformation, and generate item node embedding V'∈R^(n V ×d) to achieve effective representation of items.

7. An adaptive anchor graph attention network embedding system for large-scale sparse bipartite graphs, which is a hardware platform of any method of claims 1-6, characterized in that: It includes an adaptive anchor generation module, a message passing network module, and a bipartite graph attention network module; The adaptive anchor point generation module adaptively generates an appropriate number of anchor nodes by analyzing the scale, sparsity, number of nodes and average degree characteristics of the graph. These anchor points serve as information transfer stations to improve the efficiency of information transmission in large-scale sparse graphs. The message passing network module is used for efficient information transmission based on a fully connected attention mechanism, which enables information to be effectively transmitted between nodes of different types through dimension conversion, anchor projection, user reception and information aggregation steps; The bipartite graph attention network module processes the interaction between users and items, adopts a unified feature transformation and mask attention mechanism, and introduces an entropy regularization mechanism to effectively prevent feature degradation.

Images