A Self-Supervised Incremental Graph Node Classification Method and System Based on Adaptive Interval Playback

Abstract

This invention discloses a self-supervised incremental graph node classification method and system based on adaptive interval replay, addressing the problems of existing node classification methods such as strong dependence on supervised data, high storage and computational overhead, and poor dynamic adaptability. This invention employs self-supervised graph encoding to extract node features, eliminating the need for labeled data and improving generalization ability. It reduces storage overhead by selecting key node surrogates through progressive clustering. Based on an adaptive interval replay strategy, the replay frequency is dynamically adjusted, reducing computational costs and enhancing the retention of old knowledge. Regularized replay training is introduced to maintain the representation boundaries between tasks, ensuring that learning new tasks does not interfere with the classification of old tasks. This invention achieves efficient and low-cost unsupervised incremental node classification, applicable to dynamic graph data classification tasks such as social networks, recommendation systems, and financial risk control.

Description

Technical Field

[0001] This invention belongs to the field of graph data analysis technology, and specifically relates to a self-supervised incremental graph node classification method and system based on adaptive interval playback. Background Technology

[0002] In recent years, with the rapid development of information technology and the widespread application of graph-structured data in social networks, recommender systems, and financial risk control, node classification technology, as a core task of graph data analysis, has become a key support for improving cyberspace governance capabilities and intelligence levels. However, graph data in real-world scenarios typically exhibits dynamic growth characteristics, such as the constantly emerging new research topics in academic citation networks and the continuously updated product association networks in e-commerce platforms. This type of data often lacks annotation information and its scale expands exponentially over time, making traditional offline training methods that rely on fully labeled data unsuitable. Furthermore, frequent data updates require models to continuously adapt to new data without forgetting historical knowledge, while simple sequential training can lead to catastrophic forgetting problems, and retraining on the entire dataset faces high computational costs and storage pressures. How to achieve unsupervised, efficient, and sustainable graph node classification has become a core challenge in current research. Incremental learning offers a potential solution to the aforementioned problems. However, current incremental learning-based node classification methods face three major challenges: First, traditional supervised learning methods heavily rely on large amounts of labeled data, but in real-world scenarios, labeling is costly and difficult to obtain, limiting the model's generalization ability. Second, as the scale of graph data continues to grow, the overhead of storing and computing all node features increases dramatically, making it difficult to adapt to dynamically changing network environments. Third, the introduction of new tasks during incremental learning can easily lead to the catastrophic forgetting of knowledge from older tasks, resulting in insufficient dynamic adaptability of the model. For example, in financial risk control scenarios, fraud patterns evolve rapidly, requiring existing methods to frequently retrain the model, significantly increasing computational resource consumption.

[0003] Currently, there are relatively few research studies proposing corresponding solutions for incremental graph network scenarios. The main invention patents related to incremental learning applications include: Reference 1: Image classification method, device and equipment based on machine learning model (201911213128.5) Reference 2: Incremental learning method for deep neural network models, image processing method and device (202410880410.3) Reference 1 proposes an image classification method based on a machine learning model. This method analyzes the correlation between the current task and historical tasks, samples meta-sample data from historical tasks to generate data for online meta-learning to train the initial model, and then uses the current task data to fine-tune the model parameters.

[0004] Reference 2 proposes a deep neural network incremental learning method, which trains the model through process incremental learning and divides the incremental data into important samples and regular samples according to the predicted similarity, and optimizes the model by Dropout and normal training respectively.

[0005] The aforementioned technical solutions all rely on fully supervised learning, require labeled data, and are not adapted to topological relationship modeling of graph structure data. They employ fixed-frequency playback mechanisms, resulting in high storage and computational overhead, making them unsuitable for dynamic graph data scenarios. However, in reality, labeling dynamic graph data is costly, and new tasks often lack label support, making it difficult to directly apply supervised incremental learning methods to unlabeled wild data streams. Therefore, how to accurately complete node classification tasks using only the inherent characteristics of unlabeled data has become a pressing problem. Summary of the Invention

[0006] This invention aims to provide a self-supervised incremental graph node classification method and system based on adaptive interval replay. This invention employs self-supervised graph encoding to extract node features, eliminating the need for labeled data and improving generalization ability. It reduces storage overhead by selecting key node surrogates through progressive clustering. Based on an adaptive interval replay strategy, the replay frequency is dynamically adjusted, reducing computational costs and enhancing the retention of old knowledge. Regularized replay training is introduced to maintain the representation boundaries between tasks, ensuring that learning new tasks does not interfere with the classification of old tasks. This invention achieves efficient and low-cost unsupervised incremental node classification, suitable for dynamic graph data classification tasks such as social networks, recommendation systems, and financial risk control.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A self-supervised incremental graph node classification method based on adaptive interval playback includes: 1) Self-supervised graph coding During each task learning process, the newly added graph data is acquired and input into the self-supervised learning model for feature encoding to obtain the embedding matrix of each node in the graph data of the current task. 2) Progressive node proxy extraction The embedding matrix is ​​initially clustered and divided into multiple clusters, and the compactness of each cluster is calculated. Then, the clusters with lower compactness are recursively split until the compactness variance of all clusters is less than a set threshold. Then, the node closest to the cluster center is stored as a key node proxy in the memory buffer. 3) Adaptive interval playback task selection Based on the embedding correlation of key nodes in the memory buffer, the memory retention rate of all historical tasks is calculated, and then historical tasks with a memory retention rate lower than a set threshold are added to the current playback sequence. 4) Regularized replay training Based on the historical tasks in the current replay sequence, a replay loss function is designed, and combined with the encoding loss function of the newly added graph data, this self-supervised learning model is jointly trained. 5) Reasoning and Incremental Node Classification After the self-supervised learning model has been trained, it can be used to perform inference and classification on any graph data, including newly added nodes or previously seen old nodes.

[0008] A further improvement of this invention is that, in step 1), the specific method for self-supervised graph encoding is as follows: Step 1: Data Preprocessing Enter the current task graph data Standardization processes are performed, among which Represents a set of nodes. Represents the set of edges; Step 2: Generate Enhanced View Two enhanced views are generated using a symmetric network architecture. and The formula is expressed as:

[0009] in, This represents a data augmenter using probability. Randomly mask node attributes, probability of use Randomly shielded edge connections; Step 3: Encoding Optimization The Graph Barlow Twins model is used as the encoder to compute the embedding matrix of the augmented view. and And through the Barlow Twins loss function Optimize the encoder to generate a node embedding matrix. ,in Indicates the current task The encoder network parameters.

[0010] A further improvement of this invention is that, in step 2), the specific method for progressive node proxy extraction is as follows: Step 1: Initial Clustering graph data Embedding matrix Input the K-Means algorithm, and initially partition into: Cluster ; Step 2: Compactness Calculation For each cluster Computational compactness The formula is:

[0011] in For clusters Cluster center quantity Cluster The node embedding vector within, Cluster The number of points inside, Represents Euclidean distance; Step 3: Dynamic Cluster Splitting Calculate the mean compactness of all clusters. and variance If the cluster compactness satisfies ≥ Select the node closest to the center as the key node proxy and store it in the memory buffer. Otherwise, proportionally Further split the subclusters and recalculate and recursively repeat until the compactness variance of all subclusters satisfies the condition. ;in Representing graph data The total number of nodes in This is a preset compactness variance threshold; Step 4: Proxy caching Final extraction Each node agent stores As the current task Key knowledge representation.

[0012] A further improvement of this invention is that, in step 3), the specific method for selecting the adaptive interval playback task is as follows: Step 1: Correlation Calculation historical mission The memory buffer corresponding to the key node agent Calculate its initial embedding correlation The formula is:

[0013] in Indicates in the task The parameters of the self-supervised learning model at the end of training. This represents the matrix transpose operation. This indicates the average pooling operation. Represents the cosine similarity function; Step 2: Memory Retention Rate Monitoring Current task After learning, the relevance score sequence is dynamically updated. ,in Representative task Best memory:

[0014] Then calculate its memory retention rate. The formula is:

[0015] Step 3: Replay Task Filtering If the memory retention rate of the current historical task meets the requirements < Then the task Add to current playback sequence ;in This is the preset memory retention rate threshold.

[0016] A further improvement of this invention is that, in step 4), the specific method for regularized replay training is as follows: Step 1: Playback Loss Design The training loss for the replay task is The formula is:

[0017] Among them, the correlation preservation constraint term is:

[0018] Task difference constraints:

[0019] In the above formula, For dynamic weighting coefficients, This represents the mean squared error loss function. Indicates the current task The memory buffer, This indicates the size of the set of tasks in the current playback sequence; Step 2: Joint Training By combining the encoding loss of the newly added graph data with the memory loss of the playback task, a self-supervised learning model is trained. The overall loss function is:

[0020] in As a weighting factor, it balances the plasticity and stability of the self-supervised learning model.

[0021] A further improvement of this invention is that, in step 5), the specific method for inference and incremental node classification is as follows: Step 1: Generate node embeddings The G-BT model embeds nodes based on graph data from the current or previous tasks. Step 2: Predict Classification The node embedding representation is predicted and classified using a classifier.

[0022] A self-supervised incremental graph node classification system based on adaptive interval playback includes: Self-supervised graph encoding unit: In each task learning process, it acquires the newly added graph data and inputs it into the self-supervised learning model for feature encoding to obtain the embedding matrix of each node in the graph data of the current task; Progressive node proxy extraction unit: performs preliminary clustering on the embedding matrix and divides it into multiple clusters, thereby calculating the compactness of each cluster; then, recursively splits the clusters with lower compactness until the compactness variance of all clusters is less than a set threshold, and then stores the node closest to the cluster center in each cluster as a key node proxy in the memory buffer. Adaptive interval replay task selection unit: Based on the embedding correlation of key nodes in the memory buffer, calculate the memory retention rate of all historical tasks, and then add historical tasks with a memory retention rate lower than a set threshold to the current replay sequence; Regularized replay training unit: Design a replay loss function based on the historical tasks in the current replay sequence, and combine it with the encoding loss function of the newly added graph data to jointly train this self-supervised learning model; Inference and Incremental Node Classification Unit: After the self-supervised learning model has been trained, it can input any graph data for inference and classification, including newly added nodes or previously seen old nodes.

[0023] A further improvement of this invention lies in that, in the self-supervised graph encoding unit, the specific method of self-supervised graph encoding is as follows: Step 1: Data Preprocessing Enter the current task graph data Standardization processes are performed, among which Represents a set of nodes. Represents the set of edges; Step 2: Generate Enhanced View Two enhanced views are generated using a symmetric network architecture. and The formula is expressed as:

[0024] in, This represents a data augmenter using probability. Randomly mask node attributes, probability of use Randomly shielded edge connections; Step 3: Encoding Optimization The Graph Barlow Twins model is used as the encoder to compute the embedding matrix of the augmented view. and And through the Barlow Twins loss function Optimize the encoder to generate a node embedding matrix. ,in Indicates the current task The encoder network parameters.

[0025] A further improvement of this invention lies in the fact that the specific method for progressive node proxy extraction in the progressive node proxy extraction unit is as follows: Step 1: Initial Clustering graph data Embedding matrix Input the K-Means algorithm, and initially partition into: Cluster ; Step 2: Compactness Calculation For each cluster Computational compactness The formula is:

[0026] in For clusters Cluster center quantity Cluster The node embedding vector within, Cluster The number of points inside, Represents Euclidean distance; Step 3: Dynamic Cluster Splitting Calculate the mean compactness of all clusters. and variance If the cluster compactness satisfies ≥ Select the node closest to the center as the key node proxy and store it in the memory buffer. Otherwise, proportionally Further split the subclusters and recalculate and recursively repeat until the compactness variance of all subclusters satisfies the condition. ;in Representing graph data The total number of nodes in This is a preset compactness variance threshold; Step 4: Proxy caching Final extraction Each node agent stores As the current task Key knowledge representation.

[0027] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the self-supervised incremental graph node classification method based on adaptive interval playback.

[0028] Compared with the prior art, the present invention has at least the following beneficial technical effects: The self-supervised incremental graph node classification method and system based on adaptive interval playback provided by this invention fully explores the inherent and underlying patterns of graph data and improves existing graph node classification methods by utilizing an adaptive interval playback strategy. Compared with the prior art, the advantages of this invention are: (1) This invention introduces a self-supervised learning method for the first time in the continuous graph learning task, which breaks through the dependence on manually labeled data and enables the model to learn continuously without labels or with only a few labels, thus improving the breadth and practicality of the application.

[0029] (2) This invention proposes an adaptive interval replay strategy to intelligently and dynamically adjust the interval of data replay, reduce storage resource consumption, maximize the knowledge retention capability of the model, and improve the efficiency and performance of incremental learning.

[0030] (3) This invention combines self-supervised learning with incremental replay mechanism, enabling the model to adapt more effectively to changes in dynamic graph data, avoid catastrophic forgetting, and have stronger generalization ability in different tasks and scenarios.

[0031] (4) This invention reduces the need for storing historical data while ensuring classification accuracy, optimizes the utilization of computing resources, and improves the efficiency of training and inference, making the method more suitable for large-scale dynamic graph data scenarios.

[0032] (5) This invention is applicable to multiple fields such as social network analysis, recommendation system, financial risk control, and knowledge graph update. It can continuously learn and optimize the node classification effect in a constantly changing data environment. Attached Figure Description

[0033] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0034] Figure 1This is a flowchart of the overall framework.

[0035] Figure 2 Graph the self-supervised graph.

[0036] Figure 3 A flowchart for progressive node proxy extraction.

[0037] Figure 4 Flowchart for selecting adaptive interval playback task.

[0038] Figure 5 This is a flowchart of the regularized replay training process.

[0039] Figure 6 The flowchart for inference and incremental node classification. Detailed Implementation

[0040] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0041] In the description of this invention, it should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0042] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0043] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0044] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0045] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0046] Example 1 The self-supervised incremental graph node classification method based on adaptive interval playback provided by this invention includes: First, in each task learning phase, self-supervised graph encoding is used to extract features from newly added graph data. A graph neural network (GNN) is then used to embed node structure information, generating high-quality node representations. The model is optimized using a self-supervised objective function to capture global and local topological relationships even without labels. Second, to reduce storage overhead and improve knowledge retention, a progressive clustering method is used to adaptively select key node proxies. By gradually dividing and refining the graph data, the proxies are ensured to fully represent the graph distribution characteristics and are stored in a long-term memory buffer for subsequent replay. Third, based on the Ebbinghaus forgetting curve theory, the memory retention rate of old tasks is dynamically monitored. When the retention rate falls below a threshold, an adaptive interval replay mechanism is triggered, selecting representative node proxies from the buffer for joint training with the current task. The replay frequency is dynamically adjusted to balance knowledge integration and computational overhead. Finally, a regularized replay training strategy is introduced in incremental training. A regularized replay loss function constrains the learning of new tasks, avoiding interference with the representations of old tasks while maintaining the representation distance between tasks, reducing the risk of knowledge overwriting, and enhancing model stability and plasticity. Finally, after the model is trained, it can perform inference on any graph data, including adding or removing nodes, to achieve efficient classification.

[0047] The present invention specifically includes the following steps: 1) Self-supervised graph coding In each learning task, newly added graph data is acquired and input into a pre-trained self-supervised learning model for feature encoding. A representation learning method based on graph neural networks is employed to embed node structural information, generating high-quality node representations. During the encoding process, the model is optimized through a self-supervised objective function to improve its representational ability in the absence of labels, ensuring that the model can extract node features that contain both global structural information and local topological relationships. 2) Progressive node proxy extraction To reduce storage overhead and improve knowledge retention, a progressive clustering method is used to adaptively select key node proxies. First, the graph data is initially clustered, partitioning nodes into subsets based on similarity measures (such as Euclidean distance or cosine similarity), and calculating the compactness index of each cluster center. Then, through progressive refinement, clusters with lower compactness are recursively further subdivided to ensure that the final node proxies fully represent the distribution characteristics of the entire graph data. The selected key node proxies are stored in a long-term memory buffer for subsequent knowledge replay, mitigating the impact of catastrophic forgetting. 3) Adaptive interval playback task selection Based on the Ebbinghaus forgetting curve theory, this invention dynamically monitors the memory retention rate of older tasks to determine whether knowledge replay is necessary. Specifically, it periodically calculates the representation similarity of key node agents within the long-term memory buffer to measure their retention rate in the current model state. When the memory retention rate of certain tasks falls below a set threshold, an adaptive interval replay mechanism is triggered, selecting the most representative older task node agents and combining them with the current task for joint training. Unlike traditional continuous replay, this invention dynamically adjusts the replay frequency based on the memory retention rate, thereby effectively improving knowledge integration capabilities and reducing additional computational overhead. 4) Regularized replay training To ensure that the learning process of new tasks does not interfere with the knowledge representation of already learned tasks, a regularization constraint strategy is adopted to maintain the representation boundaries between tasks. Specifically, during incremental training, a regularized replay loss function based on the node representation space is used to ensure that the node features of the new task are learned without compromising the node distribution of the old task, while maintaining the representation distance between different tasks. This reduces the risk of knowledge from the old task being overwritten during the training of the new task. This method effectively balances the stability and plasticity of the model, enhancing its adaptability to new tasks. 5) Reasoning and Incremental Node Classification After the model is trained, any graph data can be input for inference and classification, including newly added nodes or previously seen old nodes.

[0048] In step 1) of this embodiment, the current task graph data is encoded using a self-supervised model to generate a high-quality node embedding representation, specifically including the following steps: Step 1: Data Preprocessing Enter the current task graph data Standardization processing is required; Step 2: Generate Enhanced View Two enhanced views are generated using a symmetric network architecture. and ,Right now

[0049] in, It is an enhancer that uses probability. and Randomly mask node attributes and edge connections; Step 3: Encoding Optimization The Graph Barlow Twins (G-BT) model is used as the encoder to compute the embedding matrix of the augmented view. and And through the Barlow Twins loss function Optimize the encoder.

[0050] In step 2) of this embodiment, the node embeddings are divided into multiple clusters through clustering, the compactness of each cluster is calculated, loose clusters are dynamically split and representative nodes are extracted as proxies and stored in a memory buffer. The detailed steps are as follows: Step 1: Initial Clustering Will Embedding matrix Input the K-Means algorithm, and initially partition into: Cluster ; Step 2: Compactness Calculation For each cluster Computational compactness ,in Cluster center:

[0051] Step 3: Dynamic Cluster Splitting Calculate the mean compactness of all clusters. and variance If the cluster ≥ Select the node closest to the center as the proxy and store it in the memory buffer. Otherwise, proportionally Further subdivide into subclusters, recursively calculate until all subclusters satisfy the condition. < (default = 0.1); Step 4: Proxy caching Final extraction Each node agent stores As the current task Key knowledge representation.

[0052] In step 3) of this embodiment, the memory retention rate of historical tasks is monitored, and tasks that need to be replayed are dynamically selected to ensure that important knowledge is not forgotten. This specifically includes the following steps: Step 1: Correlation Calculation historical mission node proxy Calculate its embedding correlation:

[0053] Step 2: Memory Retention Rate Monitoring Current task After learning, the relevance score sequence is dynamically updated. ,in Representative task Best memory:

[0054] Then calculate its memory retention rate. :

[0055] Step 3: Replay Task Filtering like < (default = 0.9), then the task Add to current playback sequence .

[0056] In step 4) of this embodiment, joint training is performed by combining the current task and the replay task. The specific method is as follows: Step 1: Playback Loss Design The training loss for the replay task is :

[0057] ,

[0058] in , and These are respectively relatedness preservation and task difference constraints; Step 2: Joint Training By combining the current data encoding loss with the playback task memory loss, the model is trained as follows:

[0059] in As a weighting factor, it balances the plasticity and stability of the model.

[0060] In step 5) of this embodiment, the trained model is used to perform a classification task. The specific steps are as follows: Step 1: Generate node embeddings The Graph Barlow Twins model embeds nodes based on graph data from the current or previous tasks. Step 2: Predict Classification The model predicts and classifies node embeddings.

[0061] Example 2 The CoraFull-CL dataset, containing a large number of scientific publications related to machine learning, is selected and structured into an undirected graph containing 35 tasks, each with two categories. The invention is further described in detail below with reference to the accompanying drawings, experimental examples, and specific implementation methods. All techniques implemented based on the content of this invention fall within the scope of this invention.

[0062] like Figure 1 As shown, in a specific implementation of the present invention, the self-supervised incremental graph node classification method based on adaptive interval playback includes the following steps: Step 1. Self-supervised graph coding When training on a particular task, the task data is extracted from the CoraFull-CL dataset, an augmented view is created, and then the encoder is used to encode the node representation.

[0063] Specifically, in this embodiment, the current task is input. graph data Standardization processing ( Figure 2 S101); then two enhanced views are generated using a symmetric network architecture. and Specifically, it uses probability. and Randomly mask node attributes and edge connections ( Figure 2 (S102); Finally, calculate the embedding matrix of the enhanced view. and And through the Barlow Twins loss function Optimize encoder ( Figure 2 S103).

[0064] Step 2. Progressive Node Proxy Extraction The nodes are divided into multiple clusters by clustering, the compactness of each cluster is calculated, the loose clusters are dynamically split and representative nodes are extracted as proxies and stored in the memory buffer.

[0065] Specifically, in this embodiment, initial clustering is first performed, and... Embedding matrix Input the K-Means algorithm, initially dividing the data into 10 clusters. ( Figure 3 S201); then the compactness of the initial clusters is calculated for each cluster. Computational compactness ,in Cluster center ( Figure 3 (S202); then progressive clustering is performed to calculate the mean compactness of all clusters. and variance If the cluster ≥ Select the node closest to the center as the proxy and store it in the memory buffer. Otherwise, proportionally Further subdivide into subclusters, recursively calculate until all subclusters satisfy the condition. <0.1 ( Figure 3 S203); finally extract Each node agent stores As the current task Key knowledge representation ( Figure 3 S204).

[0066] Step 3. Adaptive Interval Playback Task Selection Monitor the memory retention rate of historical tasks, dynamically select tasks that need to be replayed, and ensure that important knowledge is not forgotten.

[0067] Specifically, in this embodiment, for each historical task... The node proxy is used to calculate its embedding correlation ( Figure 4 S301); then update the relevance score for each old task. ,in Representative task Optimal memory retention and calculation of current memory retention rate. ( Figure 4 S302); Finally, based on the memory retention rate of the old task. If the value is less than the threshold of 0.9, determine whether to add it to the current playback sequence. Figure 4 S303).

[0068] Step 4. Regularized replay training By combining the current task and the replay task for joint training, the correlation between tasks is maintained and the differences between tasks are distinguished through regularization constraints, and the model parameters are updated.

[0069] Specifically, in this embodiment, the training loss of the playback task for the current playback sequence is first calculated. Figure 5S401); then, by combining the current data encoding loss and the playback task memory loss, the model is trained ( Figure 5 S402).

[0070] Step 5. Inference and Incremental Node Classification The model generates node embeddings and predicts classifications for all nodes in the input graph based on the trained network, which includes two steps: generating node embeddings and predicting classifications.

[0071] Specifically, in this embodiment, the data of each old task and the current task are input sequentially to obtain the embedded representation of each node. Figure 6 S501); then a classifier is used to classify the nodes, and the final classification result is output ( Figure 6 S502).

[0072] Example 3 The self-supervised incremental graph node classification system based on adaptive interval playback provided by this invention includes: Self-supervised graph encoding unit: In each task learning process, it acquires the newly added graph data and inputs it into the self-supervised learning model for feature encoding to obtain the embedding matrix of each node in the graph data of the current task; Progressive node proxy extraction unit: performs preliminary clustering on the embedding matrix and divides it into multiple clusters, thereby calculating the compactness of each cluster; then, recursively splits the clusters with lower compactness until the compactness variance of all clusters is less than a set threshold, and then stores the node closest to the cluster center in each cluster as a key node proxy in the memory buffer. Adaptive interval replay task selection unit: Based on the embedding correlation of key nodes in the memory buffer, calculate the memory retention rate of all historical tasks, and then add historical tasks with a memory retention rate lower than a set threshold to the current replay sequence; Regularized replay training unit: Design a replay loss function based on the historical tasks in the current replay sequence, and combine it with the encoding loss function of the newly added graph data to jointly train this self-supervised learning model; Inference and Incremental Node Classification Unit: After the self-supervised learning model has been trained, it can input any graph data for inference and classification, including newly added nodes or previously seen old nodes.

[0073] In the self-supervised graph encoding unit of this embodiment, the specific method of self-supervised graph encoding is as follows: Step 1: Data Preprocessing Enter the current task graph data Standardization processes are performed, among which Represents a set of nodes. Represents the set of edges; Step 2: Generate Enhanced View Two enhanced views are generated using a symmetric network architecture. and The formula is expressed as:

[0074] in, This represents a data augmenter using probability. Randomly mask node attributes, probability of use Randomly shielded edge connections; Step 3: Encoding Optimization The Graph Barlow Twins model is used as the encoder to compute the embedding matrix of the augmented view. and And through the Barlow Twins loss function Optimize the encoder to generate a node embedding matrix. ,in Indicates the current task The encoder network parameters.

[0075] In the progressive node proxy extraction unit of this embodiment, the specific method for progressive node proxy extraction is as follows: Step 1: Initial Clustering graph data Embedding matrix Input the K-Means algorithm, and initially partition into: Cluster ; Step 2: Compactness Calculation For each cluster Computational compactness The formula is:

[0076] in For clusters Cluster center quantity Cluster The node embedding vector within, Cluster The number of points inside, Represents Euclidean distance; Step 3: Dynamic Cluster Splitting Calculate the mean compactness of all clusters. and variance If the cluster compactness satisfies ≥ Select the node closest to the center as the key node proxy and store it in the memory buffer. Otherwise, proportionally Further split the subclusters and recalculate and recursively repeat until the compactness variance of all subclusters satisfies the condition. ;in Representing graph data The total number of nodes in This is a preset compactness variance threshold; Step 4: Proxy caching Final extraction Each node agent stores As the current task Key knowledge representation.

[0077] Example 4 The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the self-supervised incremental graph node classification method based on adaptive interval playback.

[0078] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0083] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A self-supervised incremental graph node classification method based on adaptive inter-session playback, characterized in that, include: 1) Self-supervised graph coding During each task learning process, the newly added graph data is acquired and input into the self-supervised learning model for feature encoding to obtain the embedding matrix of each node in the graph data of the current task. 2) Progressive node proxy extraction The embedding matrix is ​​initially clustered and divided into multiple clusters, and the compactness of each cluster is calculated accordingly. Subsequently, the clusters with lower compactness are recursively split until the compactness variance of all clusters is less than a set threshold. Then, the node closest to the cluster center in each cluster is stored as a key node proxy in the memory buffer. 3) Adaptive interval playback task selection Based on the embedding correlation of key nodes in the memory buffer, the memory retention rate of all historical tasks is calculated, and then historical tasks with a memory retention rate lower than a set threshold are added to the current playback sequence. 4) Regularized replay training Based on the historical tasks in the current replay sequence, a replay loss function is designed, and combined with the encoding loss function of the newly added graph data, this self-supervised learning model is jointly trained. 5) Reasoning and Incremental Node Classification After the self-supervised learning model has been trained, it can be used to perform inference and classification on any graph data, including newly added nodes or previously seen old nodes.

2. The self-supervised incremental graph node classification method based on adaptive interplay playback according to claim 1, characterized in that, In step 1), the specific method for self-supervised graph encoding is as follows: Step 1: Data Preprocessing Enter the current task graph data Standardization processes are performed, among which Represents a set of nodes. Represents the set of edges; Step 2: Generate Enhanced View Two enhanced views are generated using a symmetric network architecture. and The formula is expressed as: in, This represents a data augmenter using probability. Randomly mask node attributes, probability of use Randomly shielded edge connections; Step 3: Encoding Optimization The Graph Barlow Twins model is used as the encoder to compute the embedding matrix of the augmented view. and And through the Barlow Twins loss function Optimize the encoder to generate a node embedding matrix. ,in Indicates the current task The encoder network parameters.

3. The self-supervised incremental graph node classification method based on adaptive interval playback according to claim 2, characterized in that, In step 2), the specific method for progressive node proxy extraction is as follows: Step 1: Initial Clustering graph data Embedding matrix Input the K-Means algorithm, and initially partition into: Cluster ; Step 2: Compactness Calculation For each cluster Computational compactness The formula is: in For clusters Cluster center quantity Cluster The node embedding vector within, Cluster The number of points inside, Represents Euclidean distance; Step 3: Dynamic Cluster Splitting Calculate the mean compactness of all clusters. and variance If the cluster compactness satisfies ≥ Select the node closest to the center as the key node proxy and store it in the memory buffer. Otherwise, proportionally Further split the subclusters and recalculate and recursively repeat until the compactness variance of all subclusters satisfies the condition. ;in Representing graph data The total number of nodes in This is a preset compactness variance threshold; Step 4: Proxy caching Final extraction Each node agent stores As the current task Key knowledge representation.

4. The self-supervised incremental graph node classification method based on adaptive interval playback according to claim 3, characterized in that, In step 3), the specific method for selecting the adaptive interval playback task is as follows: Step 1: Correlation Calculation historical mission The memory buffer corresponding to the key node agent Calculate its initial embedding correlation The formula is: in Indicates in the task The parameters of the self-supervised learning model at the end of training. This represents the matrix transpose operation. This indicates the average pooling operation. Represents the cosine similarity function; Step 2: Memory Retention Rate Monitoring Current task After learning, the relevance score sequence is dynamically updated. ,in Representative task Best memory: Then calculate its memory retention rate. The formula is: Step 3: Replay Task Filtering If the memory retention rate of the current historical task meets the requirements < Then the task Add to the current playback sequence ;in This is the preset memory retention rate threshold.

5. The self-supervised incremental graph node classification method based on adaptive interval playback according to claim 4, characterized in that, In step 4), the specific method for regularized replay training is as follows: Step 1: Playback Loss Design The training loss for the replay task is The formula is: Among them, the correlation preservation constraint term is: Task difference constraints: In the above formula, For dynamic weighting coefficients, This represents the mean squared error loss function. Indicates the current task The memory buffer, This indicates the size of the set of tasks in the current playback sequence; Step 2: Joint Training By combining the encoding loss of the newly added graph data with the memory loss of the playback task, a self-supervised learning model is trained. The overall loss function is: in This is a weighting factor, ranging from 0 to 1, used to balance the plasticity and stability of the self-supervised learning model.

6. The self-supervised incremental graph node classification method based on adaptive interval playback according to claim 1, characterized in that, In step 5), the specific methods for inference and incremental node classification are as follows: Step 1: Generate node embeddings The G-BT model embeds nodes based on graph data from the current or previous tasks. Step 2: Predict Classification The node embedding representation is predicted and classified using a classifier.

7. A self-supervised incremental graph node classification system based on adaptive interval playback, characterized in that, include: Self-supervised graph encoding unit: In each task learning process, it acquires the newly added graph data and inputs it into the self-supervised learning model for feature encoding to obtain the embedding matrix of each node in the graph data of the current task; Progressive node proxy extraction unit: performs preliminary clustering on the embedding matrix and divides it into multiple clusters, thereby calculating the compactness of each cluster; Subsequently, the clusters with lower compactness are recursively split until the compactness variance of all clusters is less than a set threshold. Then, the node closest to the cluster center in each cluster is stored as a key node proxy in the memory buffer. Adaptive interval replay task selection unit: Based on the embedding correlation of key nodes in the memory buffer, calculate the memory retention rate of all historical tasks, and then add historical tasks with a memory retention rate lower than a set threshold to the current replay sequence; Regularized replay training unit: Design a replay loss function based on the historical tasks in the current replay sequence, and combine it with the encoding loss function of the newly added graph data to jointly train this self-supervised learning model; Inference and Incremental Node Classification Unit: After the self-supervised learning model has been trained, it can input any graph data for inference and classification, including newly added nodes or previously seen old nodes.

8. The self-supervised incremental graph node classification system based on adaptive interval playback according to claim 7, characterized in that, In the self-supervised graph coding unit, the specific method of self-supervised graph coding is as follows: Step 1: Data Preprocessing Enter the current task graph data Standardization processes are performed, among which Represents a set of nodes. Represents the set of edges; Step 2: Generate Enhanced View Two enhanced views are generated using a symmetric network architecture. and The formula is expressed as: in, This represents a data augmenter using probability. Randomly mask node attributes, probability of use Randomly shielded edge connections; Step 3: Encoding Optimization The Graph Barlow Twins model is used as the encoder to compute the embedding matrix of the augmented view. and And through the Barlow Twins loss function Optimize the encoder to generate a node embedding matrix. ,in Indicates the current task The encoder network parameters.

9. The self-supervised incremental graph node classification system based on adaptive interval playback according to claim 8, characterized in that, In the progressive node proxy extraction unit, the specific method for progressive node proxy extraction is as follows: Step 1: Initial Clustering graph data Embedding matrix Input the K-Means algorithm, and initially partition into: Cluster ; Step 2: Compactness Calculation For each cluster Computational compactness The formula is: in For clusters Cluster center quantity Cluster The node embedding vector within, Cluster The number of points inside, Represents Euclidean distance; Step 3: Dynamic Cluster Splitting Calculate the mean compactness of all clusters. and variance If the cluster compactness satisfies ≥ Select the node closest to the center as the key node proxy and store it in the memory buffer. Otherwise, proportionally Further split the subclusters and recalculate and recursively repeat until the compactness variance of all subclusters satisfies the condition. ;in Representing graph data The total number of nodes in This is a preset compactness variance threshold; Step 4: Proxy caching Final extraction Each node agent stores As the current task Key knowledge representation.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the self-supervised incremental graph node classification method based on adaptive interval playback as described in any one of claims 1-6.

Images