A scalable attribute missing map clustering method by neighborhood differential

Abstract

This invention discloses a scalable clustering method for attribute-missing graphs using neighborhood differentiation, belonging to the field of graph data processing technology. The method first obtains the complete multi-hop neighborhoods of nodes through recursive neighborhood search; then, it performs a difference operation on the neighborhood set using a neighborhood differentiation strategy to obtain a series of disjoint difference neighborhoods, eliminating information redundancy and extracting pure structural information; finally, it aggregates the difference neighborhoods to generate multi-view feature representations. As a plug-and-play preprocessing module, this method can directly enhance existing graph clustering or multi-view clustering algorithms, effectively solving the problem of decreased clustering performance caused by information redundancy, structural damage, or computational non-scalability in large-scale, high-attribute-missing scenarios. It has the advantages of complete and non-redundant information, efficient and scalable processing, a general framework, and strong robustness.

Description

Technical Field

[0001] This invention relates to a scalable clustering method for attribute-missing graphs using neighborhood differentiation, and more particularly to an automatic clustering method for massive Ego human-computer interaction datasets with attribute-missing graphs using neighborhood differentiation. This method can be used to automatically cluster nodes in attribute-missing scenarios of massive Ego human-computer interaction relationship graphs automatically constructed by embodied acquisition and first-person perspective acquisition projects. Background Technology

[0002] Graph clustering, especially unsupervised node clustering in attribute-missing scenarios, is a core task in graph data analysis and has widespread and urgent demand in many practical application areas. For example:

[0003] 1. Social Network Analysis (Community Detection): On platforms such as Weibo and Xiaohongshu, user nodes often lack attributes such as personal profiles. It is necessary to identify potential groups based on the interaction relationships between users (such as following and liking) and to quickly cluster and recommend new users.

[0004] 2. E-commerce and recommendation systems: For anonymous users with sparse attributes, clustering is performed using their interaction graphs with products (purchase, browsing) to generate personalized recommendations; or in the product co-purchase graph, missing descriptive attributes are supplemented through product associations to optimize category division.

[0005] 3. Bioinformatics and Healthcare: In protein-protein interaction networks (PPIs), some proteins lack functional annotations and their functional modules need to be inferred through multi-hop neighborhood structures; based on patient gene expression networks, subtype classification of patients lacking clinical labels can be performed to assist in precision medicine.

[0006] 4. Network security and anomaly detection: Malicious accounts often forge attributes, and clustering based on their connection patterns (such as frequent creation of abnormal edges) is necessary to identify suspicious groups; in the device communication graph, abnormal behavior clustering is performed on devices lacking log attributes to identify attack sources.

[0007] 5. Urban Traffic and IoT: When sensor data is missing at traffic intersections, cluster intersections with similar traffic patterns using the traffic flow correlation of multi-hop neighborhoods; in IoT device networks, supplement missing attributes and optimize collaborative strategies through device interaction relationships.

[0008] 6. Academic Research and Knowledge Graphs: In citation networks, papers with missing abstracts are clustered by topic through citation relationships; in knowledge graphs, entities with missing attributes (such as historical figures) are inferred by multi-hop association entities.

[0009] 7. Embodied Intelligence and First-Person Perspective Interaction Data: Embodied data collection naturally generates large-scale interaction graphs with missing attributes, which is the core application scenario of this method. Taking first-person (ego-centric) data collection projects such as EgoDex and EgoExo4D as examples, they automatically construct massive interaction relationship graphs during the recording of interactions between human hands and objects, and between humans and the environment. However, due to limitations in data collection conditions, the semantic attributes of many nodes in the graph (such as occluded objects and instantaneous interaction objects) are often missing. It is necessary to use the interaction structure of multi-hop neighborhoods to automatically cluster them to support downstream tasks such as action pattern discovery and interaction object classification.

[0010] To address the graph clustering tasks in the aforementioned applications, existing technologies have mainly developed the following types of technical solutions:

[0011] 1. The "aggregation-encoding" paradigm of traditional deep graph clustering (DGC): This paradigm is the foundation of most graph neural networks (GNNs) and deep graph clustering methods. Its core is to iteratively aggregate neighbor information to update the representation of the central node through a message propagation mechanism. For example, in a graph convolutional network (GCN), the k-th order representation of a node... It is achieved by aggregating the k-1 order representations of its neighbors ( This is obtained from... For example, a typical GCN layer can be represented as: Where à is the normalized adjacency matrix, W is the learnable weight matrix, and σ is the activation function. This process is repeated K times, and the node obtains information about its K-hop neighbors. This process is repeated K times to fuse the K-hop neighbor information.

[0012] 2. Scalable solutions for large-scale graphs: To address the high computational cost and memory consumption of traditional DGC, the mainstream approach is based on sampling techniques. These methods do not directly compute the entire graph, but instead sample one or more subgraphs in each training batch for GNN computation. Representative methods include GraphSAGE (neighbor sampling), Cluster-GCN (subgraph sampling), and Dink-Net (random walk sampling).

[0013] 3. Solutions for graphs with missing attributes: These techniques specifically address the problem of missing node features and are mainly divided into two categories:

[0014] Feature imputation / propagation (e.g., FeaturePropagation, FP): As a preprocessing step, missing features are "filled" with neighboring nodes of known features, and the process is iterated on the graph in a manner similar to label propagation until convergence.

[0015] End-to-end models (such as AMGC): These are models designed to directly handle missing values. They are optimized by iteratively alternating between clustering objectives and attribute imputation tasks. They are currently the only end-to-end models specifically designed for clustering graphs with missing attributes.

[0016] 4. Multi-view graph representation learning scheme: The core idea is to construct multiple "views" from the graph and learn node representations by maximizing the consistency between different views through comparative learning. Representative works such as MVCRL construct a structural view through graph structure enhancement (as shown in graph diffusion), and form a multi-view with the original feature view. Comparative learning brings the representation of the same node closer together in different views.

[0017] While the above-mentioned solutions have promoted the development of graph clustering technology at different levels, they all have significant shortcomings when dealing with complex real-world scenarios where "large scale" and "attribute loss" coexist:

[0018] 1. Shortcomings of the traditional depth map clustering (DGC) paradigm:

[0019] Information redundancy and oversmoothing: Iterative message propagation based on GNNs leads to excessive weighting and repeated computation of nearest neighbor information, while diluting the unique information of distant neighbors. After multi-layer aggregation, node representations tend to be consistent, losing their distinguishability and severely impairing clustering performance, especially in scenarios where attributes that rely on structural information are missing.

[0020] High computational cost and lack of scalability: Full graph GNN propagation involves huge matrix operations, and the computational complexity and memory consumption increase dramatically with the size of the graph, making it prone to memory overflow when processing large graphs with millions of nodes, thus making it unsuitable for practical applications.

[0021] 2. Limitations of scalable graph clustering schemes:

[0022] Loss of key structural information: Sampling techniques (such as node sampling and subgraph sampling) discard some graph information to reduce computation, resulting in incomplete subgraphs used for computation, and many original edges and neighbor relationships are cut off. In scenarios with missing attributes, the graph's topology is a key source of information for clustering. Sampling destroys the integrity of this information, severely weakening the model's ability to infer node relationships from the structure, leading to a sharp decline in performance.

[0023] 3. Deficiencies of the missing attribute graph clustering scheme:

[0024] Unable to scale to large-scale graphs: Dedicated methods, such as AMGC, still use the standard GCN as their internal encoder. Therefore, they completely inherit the lack of scalability of traditional DGC and can only run on small graphs. They cannot cope with the dual challenges of large-scale graphs with missing attributes.

[0025] 4. Limitations of the multi-view graph learning scheme:

[0026] Randomness and information opacity in view construction: Traditional methods generate views through random data augmentation (such as random edge dropping) or complex graph transformations (such as graph diffusion). The resulting views have high information overlap, and the unique information represented by each view is opaque and uninterpretable. They rely on contrastive learning to "expect" the model to learn complementary information, but this is not guaranteed mechanistically, leading to low learning efficiency and suboptimal results.

[0027] In summary, existing technologies have failed to effectively address the core challenge of fully and without redundancy utilizing graph structure information and achieving efficient, scalable, and accurate clustering in complex scenarios with large-scale missing node attributes and massive graph size. Specifically, traditional propagation-based methods suffer from information redundancy and computational bottlenecks; scalable sampling methods sacrifice key structural information; methods specializing in missing attributes cannot be scaled; and multi-view methods lack a mechanism to ensure view complementarity. Therefore, a novel technical solution is urgently needed to simultaneously overcome the triple challenges of information integrity, computational scalability, and robustness to missing attributes. Summary of the Invention

[0028] To address the shortcomings of existing technologies, this invention discloses a scalable attribute-missing graph clustering method based on neighborhood differentiation. The technical solution is as follows: A graph information processing method based on neighborhood differentiation (hereinafter referred to as "neighborhood differentiation-based multi-view graph clustering method," abbreviated as CMV-ND), comprising the following steps:

[0029] Step 1: For each node in the graph Recursively search and determine its jump from the 1st to the 2nd. The neighborhood set of jumps, To preset the maximum number of jumps, The value range is 2 to 5; where, the defined node... The 0th hop neighborhood is itself. ,node The Jump Neighborhood From its first Jump Neighborhood as well as It consists of the direct neighbor nodes of all nodes in the array, i.e. ,in Let the edge set of the graph form a series of nested neighborhood sets. ;

[0030] Step 2: For each , will node of Jump Neighborhood With Jump Neighborhood Perform set difference operation to generate -Difference Neighborhood ,Right now The Only includes distance nodes Exactly Jump node;

[0031] Step 3: For each node Constructing multi-view feature representations includes: [representing nodes] Its own original feature vector As the 0th view representation If node If the original features are missing, then... Initialize as zero vector For each ,right -Difference Neighborhood Feature vectors of all nodes Perform aggregation operation to generate the first... View representation ;Will By splicing together, nodes are formed. Multi-view feature representation .

[0032] Preferably, in step one, a breadth-first search algorithm is used to recursively and exhaustively explore and record the maximum number of hops for each node. All neighbors within the range to ensure that local topology information is fully captured.

[0033] Preferably: in step three, for each -Difference Neighborhood The aggregation operation is as follows: calculate all node feature vectors The arithmetic mean of the first digits is used to generate the second digit. View representation ,in Represents a set The number of elements in the middle; if If the set is empty, then a zero vector is generated. This is represented by the view.

[0034] Preferably, steps one and two are one-time deterministic preprocessing processes executed on the central processing unit and do not involve message propagation or iterative matrix operations in the graph neural network, thereby avoiding memory overflow when processing large-scale graphs and achieving scalability.

[0035] This invention also discloses a scalable attribute-missing graph clustering method based on neighborhood differentiation, comprising:

[0036] Using the method described above, a multi-view feature representation is generated for each node in the graph to be clustered. ; Represent the multi-view features The data is input to the downstream clustering algorithm to perform node clustering.

[0037] This invention also discloses a graph data processing system based on neighborhood differentiation, which performs the above-described method, characterized in that it includes:

[0038] The recursive neighborhood search module is configured to perform the operation described in step one;

[0039] The neighborhood differentiation processing module is configured to perform the operation described in step two;

[0040] The multi-view building module is configured to perform the operation described in step three.

[0041] Preferably, the recursive neighborhood search module and the neighborhood differentiation processing module run on a central processing unit, supporting parallel processing of large-scale graphs.

[0042] This invention also discloses a graph clustering system, characterized in that it includes: using the above-described graph data processing system to generate multi-view feature representations of nodes in the graph. Clustering execution module, used to receive the multi-view feature representation. Then run the clustering algorithm to output the node clustering results.

[0043] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the above-described method.

[0044] The present invention also discloses a terminal device, characterized in that the terminal device includes: a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected through the bus and communicate with each other; the memory stores executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to execute the above-mentioned method.

[0045] Beneficial effects

[0046] (1) This invention combines "recursive neighborhood search + neighborhood differentiation strategy" and uses set difference operation, a deterministic algebraic operation, to replace the iterative message propagation that relies on learnable parameters in traditional GNNs. When faced with the problem of information redundancy, those skilled in the art usually design more complex attention mechanisms or regularization methods (such as Dropout), rather than using set difference to explicitly decompose the neighborhood. This invention innovatively introduces set difference operation into graph neighborhood processing, achieving redundancy-free decomposition of neighborhood information in a zero-parameter, zero-training manner. This technical path is not recorded in existing graph clustering literature and has non-obviousness.

[0047] (2) By adopting a non-propagational, deterministic preprocessing paradigm, the core computations (neighborhood search and difference operations) are executed in parallel on the CPU, eliminating the need for expensive, iterative GNN full-graph matrix operations. This completely solves the memory overflow problem of traditional deep graph clustering and AMGC methods, and for the first time makes it possible to perform attribute-missing clustering on large-scale graphs with millions of nodes, filling the technical gap in this complex scenario.

[0048] (3) The entire process is designed as an independent preprocessing module, outputting a standardized multi-view feature representation. This representation can be seamlessly integrated into various downstream graph clustering or multi-view clustering algorithms without modifying the main model. It can enhance the performance of existing DGC models (such as DGI and Dink-Net), and also bridge the fields of graph clustering and multi-view clustering. It utilizes the mature MVC algorithm to process graph data, and has high versatility and ease of use.

[0049] (4) Even if a node's own attributes are completely missing, its representation can still be constructed from a structured view obtained by aggregating multiple differential neighborhoods, thereby effectively compensating using the topological context. Under extreme conditions with extremely high node attribute missing rates (e.g., missing rate of 90%), the present invention can still maintain stable clustering performance, overcoming the shortcomings of existing methods that suffer from a sharp drop in performance when the data is severely sparse, and expanding the practical boundaries of the method in harsh data environments.

[0050] (5) The innovative process of “complete capture → redundancy-free decomposition → multi-view construction” replaces the traditional “propagation-aggregation” coding paradigm. It not only comprehensively solves the contradiction between performance, scale and missing robustness, but also provides a new technical path for the field of graph representation learning with its preprocessing ideas, which has important theoretical value and broad industrial application potential. Attached Figure Description

[0051] Figure 1 This is an overall flowchart of the method of the present invention (a multi-view graph clustering method based on neighborhood differentiation, abbreviated as CMV-ND);

[0052] Figure 2This is a schematic diagram of recursive neighborhood search (RNS).

[0053] Figure 3 A schematic diagram of the Neighborhood Differentiation Strategy (NDS);

[0054] Figure 4 Construct a schematic diagram for multi-view feature representation;

[0055] Figure 5 This is a flowchart of the method of the present invention. Detailed Implementation

[0056] like Figure 1 As shown, the overall process of the method of the present invention includes three core stages: recursive neighborhood search (RNS), neighborhood differentiation strategy (NDS), and multi-view feature representation construction. The following describes each step in detail with reference to the accompanying drawings.

[0057] A graph information processing method based on neighborhood differentiation includes the following steps:

[0058] Step 1: Recursive Neighborhood Search (RNS)

[0059] For each node in the graph Recursively search and determine its jump from the 1st to the 2nd. The neighborhood set of jumps, The maximum number of hops is preset. This step is the initial stage of the graph information processing method, and its primary goal is to fully capture the multi-level topological structure of the local neighborhood of each node in the graph. This goal is achieved through a preprocessing program running on a central processing unit (CPU), whose processing object is the adjacency list of the original graph. Its core mechanism is: for each node in the graph... This triggers a deterministic, breadth-first search-like recursive exploration process in sequence.

[0060] Specifically, the process begins with setting up nodes It is defined as its 0-hop neighborhood, that is Subsequently, the program... Based on the diagram Edge connection relationship It begins to expand outwards layer by layer. The expansion follows a strict recursive definition: the first... Jump Neighborhood By the Jump Neighborhood All nodes in the set, plus all their direct neighbors, constitute the set, which is expressed as: This operation is executed recursively, as if using nodes. Using the center as the pivot, draw a series of concentric circles with ever-increasing radii.

[0061] User-preset maximum number of exploration jumps These are the key operating parameters for this step. In a preferred embodiment, The value range is 2 to 5, recommended. For most social networks and citation networks, a 3-hop neighborhood is sufficient to cover the main structural information of nodes, and the computational cost is manageable; for large-scale graphs with more than 1 million nodes, it is recommended to... This reduces computational complexity. Those skilled in the art can reasonably choose a value within the range of 2 to 5 based on the specific size and density of the graph. value. As the termination condition of the process, it precisely defines the spatial scope of information capture. Its operating principle lies in ensuring, through an exhaustive, layer-by-layer expanding search, that within... All nodes within the reachable range of the jump path are considered without exception. When the search depth reaches... The process ends at that time.

[0062] Unlike existing technologies that rely on iterative, implicit message propagation using graph neural networks (GNNs), this step employs an explicit, non-propagating, one-time structure exploration method. It directly operates on the graph's basic topology (adjacency list), explicitly enumerating and recording physical connections using a deterministic broad-first search (BFS) algorithm. This avoids the information redundancy and long-distance information dilution problems caused by multi-layer aggregation in the GNN paradigm. This design not only guarantees 100% complete acquisition of local structural information but also ensures predictable and controllable computational overhead, laying a scalable foundation for subsequent processing of large-scale graphs.

[0063] Finally, for each node This step outputs a series of nested neighborhood sets: The result is visually presented as a representation around the node. Multiple concentric regions, where the inner region is completely contained within the outer region. For example, Includes and all its direct neighbors; In containing While including all nodes, the algorithm further expands to include the neighbors of these direct neighbors; and so on. Although this nested set is complete, it contains significant information overlap (redundancy), which is the core problem that subsequent steps need to address. The implementation of this step provides the entire method with a raw, complete, and unfuzzy local structure data foundation.

[0064] Step 2: Neighborhood Differentiation Strategy (NDS)

[0065] For each , will node of Jump Neighborhood With Jump Neighborhood Perform set difference operation to generate -Difference Neighborhood ,Right now The Only includes distance nodes Exactly Jump nodes.

[0066] This step aims to perform a redundancy-free decomposition on the nested neighborhood set generated in Step 1, which contains significant information redundancy, in order to extract pure, incremental local structural information. This step also functions as a preprocessing program running on the CPU, processing the nested neighborhood set output from Step 1. The trigger condition is that step one completes the recursive neighborhood search (RNS) for all nodes.

[0067] The specific work process is as follows: For each node The program processes each of its sequentially. From 1 to The neighborhood set of each pair of adjacent levels. For each level Perform a specific set difference operation: The meaning of this operation is that, from Jump Neighborhood Remove all items already included in the middle. Jump Neighborhood The nodes in the process. This process is repeated until all nodes are processed. layer.

[0068] Unlike existing Graph Neural Networks (GNNs), which implicitly mix information from neighbors at different distances through weighted aggregation, leading to information ambiguity and redundancy, this step employs an explicit, algebraic "differentiation" operation. It discretizes and differs continuous, overlapping neighborhood layers, thereby separating a unique set of newly added nodes at each specific distance level. This step fundamentally eliminates the information redundancy and circular reference problems inherent in traditional message propagation paradigms. Both precisely represent the distance of... The "information ring" formed by the hop nodes clearly separates and independently preserves the topological contributions from different network distances. This lays the mathematical foundation for the subsequent construction of complementary, non-redundant multi-view representations.

[0069] When for all This step ends after the difference operation is completed. Finally, for each node... We obtain a series of disjoint difference neighborhoods: .like Figure 3 As shown, Corresponding to the ring formed by direct neighbors, A ring corresponding to exactly two hops of neighbors, and so on. These ring-shaped regions collectively and without overlap cover the nodes. exist The complete local structure within each hop, with each ring carrying a layer of pure topological information from a specific network distance. The output of this step achieves the crucial transformation of the node's local structure from "complete but redundant" to "complete and non-redundant".

[0070] Step 3: Construction of Multi-View Feature Representation

[0071] For each node Constructing multi-view feature representations includes: [representing nodes] Its own original feature vector As the 0th view representation For each ,right -Difference Neighborhood Feature vectors of all nodes Perform aggregation operation to generate the first... View representation ;Will By splicing together, nodes are formed. Multi-view feature representation .

[0072] This step is the integration and output stage of the method. Its core objective is to fuse the non-redundant pure structural information extracted in step two with the original attribute information of the nodes, constructing a unified, complementary, and directly applicable multi-view feature representation for downstream machine learning tasks. This step is also executed by the CPU preprocessor, and its processing object includes the set of disjoint difference neighborhoods output in step two. and the original node feature matrix (in Represents a node eigenvectors, Represents a node The feature vector). Its triggering condition is that the preceding neighborhood differentiation strategy (NDS) has finished processing all nodes.

[0073] The specific work process follows a standardized build process:

[0074] (1) View generation:

[0075] • View 0 (self-view): Directly extract nodes Its own original feature vector is defined as It should be particularly noted that, regarding the "attribute missing" scenario that this invention addresses as its core problem, if the node... The original features were completely missing during the data acquisition phase (i.e.) If the value is undefined or all zeros, this step initializes it as a zero vector. At this point, although the 0th view does not contain valid attribute information, the subsequent multi-scale views generated by the neighborhood structure will effectively compensate for it. This view carries the inherent attribute information of the nodes.

[0076] ·No. View (Structure View): For each For differential neighborhood The set of nodes in the dataset is used to perform a feature aggregation operation. In this embodiment, mean aggregation is used. This operation condenses the features of discrete neighboring nodes belonging to the same "ring" into a generalized continuous vector. As a robust design, if... If the distance layer is empty (i.e., it has no neighbors), then a zero vector is generated. This is represented by the view.

[0077] (2) Representation integration: Concatenate all generated view representations in sequence to form nodes. The final comprehensive representation: .

[0078] To illustrate this more clearly, let's take a concrete example: Suppose a node... The original feature vector (dimension) ), its 1-difference neighborhood 2-Difference Neighborhood Then, after aggregating the means... The dimension is 3. (Dimension 3). The final multi-view feature representation obtained by stitching together. The dimension is Wei. If If it is an empty set, then (zero vector) If node If its own attributes are missing, then ,at this time ,but and It still provides valid structural context information.

[0079] This step implements a deterministic multi-view construction mechanism based on explicit structural semantics. Unlike existing multi-view learning methods (such as MVGRL) that generate semantically ambiguous and information-overlapping views through random graph augmentation or complex diffusion, each structured view in this invention... Each of these uniquely and explicitly corresponds to a specific distance level in the network topology. (Jump). Due to the underlying differential neighborhood The non-overlapping nature of these views ensures complementarity and non-redundancy at the information source. View 0 provides a "self" description, while view... Figure 1 arrive This provides a pure "structural context" description from different scales, from near to far, which together constitute a panoramic picture of the node in the multi-scale topological space.

[0080] Application instructions for downstream clustering algorithms

[0081] This step concludes after the above construction is completed for all nodes in the graph. The final output is a graph with dimension [missing information]. The three-dimensional tensor (where The total number of nodes. (This refers to the original feature dimension), which is the set of plug-and-play multi-view feature representations for all nodes. For example... Figure 4 As shown, this representation integrates its own attributes with a multi-layered topological context.

[0082] In practical graph clustering tasks, this feature represents a set. It can be directly input into various general-purpose clustering algorithms. For example, it can be reshaped into... The two-dimensional feature matrix is ​​used, and then the K-means clustering algorithm is applied to iteratively optimize the sum of squared intra-cluster distances to partition nodes. Alternatively, the cosine similarity or Euclidean distance between nodes can be calculated based on this feature matrix to construct a similarity matrix, which is then input into the Spectral Clustering algorithm for clustering. Because this feature representation itself is highly abstract and retains complete, non-redundant structural information, even using the simplest K-means algorithm, it can achieve clustering accuracy far exceeding traditional graph neural network methods on large-scale graphs with highly missing attributes.

[0083] This invention addresses the core technical shortcomings of existing technologies in processing large-scale, high-attribute-missing graph data, including information redundancy, computational non-scalability, structural information destruction, and poor model universality. It proposes a novel, non-propagating paradigm for encoding and representing graph structural information. This paradigm captures the local topology of nodes through recursive neighborhood search, performs deterministic set difference operations on the neighborhood set using a neighborhood differentiation strategy to eliminate information redundancy, extract pure incremental structural information, and ultimately constructs a multi-view feature representation based on explicit structural semantics. This series of collaborative techniques achieves, for the first time, complete and redundant structural information extraction and encoding of large-scale graphs with millions of nodes without sampling or relying on iterative message propagation. The resulting plug-and-play multi-view representation not only significantly improves the accuracy and robustness of existing graph clustering and multi-view clustering algorithms in attribute-missing scenarios but also fundamentally overcomes the scalability bottleneck of traditional methods. It provides an efficient, universal, and reliable underlying solution for many fields requiring the processing of complex and incomplete graph data, such as social network analysis, recommendation systems, bioinformatics, and network security.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A graph information processing method based on neighborhood differentiation, characterized in that, Includes the following steps: Step 1: For each node in the graph Recursively search and determine its jump from the 1st to the 2nd. The neighborhood set of jumps, To preset the maximum number of jumps, The value range is 2 to 5; where, the defined node... The 0th hop neighborhood is itself. ,node The Jump Neighborhood From its first Jump Neighborhood as well as It consists of the direct neighbor nodes of all nodes in the array, i.e. ,in Let the edge set of the graph form a series of nested neighborhood sets. ; Step 2: For each , will node of Jump Neighborhood With Jump Neighborhood Perform set difference operation to generate -Difference Neighborhood ,Right now The Only includes distance nodes Exactly Jump node; Step 3: For each node Constructing multi-view feature representations includes: [representing nodes] Its own original feature vector As the 0th view representation If node If the original features are missing, then... Initialize as zero vector For each ,right -Difference Neighborhood Feature vectors of all nodes Perform aggregation operation to generate the first... View representation ;Will By splicing together, nodes are formed. Multi-view feature representation .

2. The method according to claim 1, characterized in that, In step one, a breadth-first search algorithm is used to recursively and exhaustively explore and record the maximum number of hops for each node. All neighbors within the range to ensure that local topology information is fully captured.

3. The method according to claim 1, characterized in that, In step three, for each -Difference Neighborhood The aggregation operation is as follows: calculate all node feature vectors The arithmetic mean of the first digits is used to generate the second digit. View representation ,in Represents a set The number of elements in the middle; if If the set is empty, then a zero vector is generated. This is represented by the view.

4. The method according to claim 1, characterized in that, Steps one and two are one-time deterministic preprocessing processes executed on the central processing unit and do not involve message propagation or iterative matrix operations in graph neural networks, thereby avoiding memory overflow when processing large-scale graphs and achieving scalability.

5. A scalable clustering method for attribute-missing graphs using neighborhood differentiation, characterized in that, include: Using the method described in any one of claims 1 to 4, generate a multi-view feature representation for each node in the graph to be clustered. ; Represent the multi-view features The data is input to the downstream clustering algorithm to perform node clustering.

6. A graph data processing system based on neighborhood differentiation, wherein the system performs the method of claim 1, characterized in that, include: The recursive neighborhood search module is configured to perform the operation described in step one; The neighborhood differentiation processing module is configured to perform the operation described in step two; The multi-view building module is configured to perform the operation described in step three.

7. The system according to claim 6, characterized in that, The recursive neighborhood search module and the neighborhood differentiation processing module run on the central processing unit, supporting parallel processing of large-scale graphs.

8. A graph clustering system, characterized in that, include: The graph data processing system as described in claim 6 or 7 is used to generate multi-view feature representations of nodes in a graph. Clustering execution module, used to receive the multi-view feature representation. Then run the clustering algorithm to output the node clustering results.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to perform the method described in any one of claims 1 to 4.

10. A terminal device, characterized in that, The terminal device includes: a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected through the bus and communicate with each other; the memory stores executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to perform the method of any one of claims 1-4.

Images