Node Classification Model Training via Target Node Subset Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graph Convolutional Networks (GCNs) face high computational overhead and resource consumption when traversing every node in large-scale graphs for feature information updates, leading to inefficient node classification processes.
Innovation Solution
A method for training a node classification model that selects a target node subset and its associated neighbor nodes, extracts feature information using a node classification model, performs class prediction, and trains the model using predicted class probabilities, thereby reducing the need to traverse all nodes in each iterative calculation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If GCN traverses every node in large-scale graphs for feature information updates, then comprehensive node classification accuracy is improved, but computational overhead and resource consumption become excessively high
Solution Approach 1:
The patent segments the graph nodes into different layers based on their distance from seed nodes. Instead of processing all nodes uniformly, the method divides nodes into first-layer nodes (direct neighbors of seed nodes) and second-layer nodes (neighbors of first-layer nodes). This segmentation allows the system to focus computational resources on the most relevant nodes, reducing the overall computational burden while maintaining classification accuracy for the nodes of interest.
Solution Approach 2:
The patent applies local quality by treating different nodes with different levels of processing detail based on their importance. Seed nodes and their immediate neighbors (first-layer nodes) receive full attention and detailed feature extraction, while other nodes are either processed with reduced detail or not processed at all. This localized approach ensures high accuracy for critical nodes while reducing unnecessary computation for less important nodes.
2Loss of information
If GCN processes all nodes in each iterative calculation, then complete graph information is captured, but computing resources are excessively consumed
Solution Approach 1:
The patent extracts and processes only the most relevant subset of graph information - specifically, seed nodes, their direct neighbors (first-layer nodes), and neighbors of those nodes (second-layer nodes). By extracting this localized subgraph, the method captures the essential graph information needed for classification while discarding redundant information from distant nodes, thereby reducing computing resource consumption without significant loss of critical graph structure.
Solution Approach 2:
The patent applies partial action by processing only a portion of the graph nodes that are most relevant to the classification task. Instead of performing exhaustive processing on all nodes, the method focuses computational effort on seed nodes and their local neighborhoods, which are sufficient for achieving accurate classification results while consuming fewer computing resources.
3Measurement precision
If feature information of each node is updated in every iteration, then classification accuracy is maintained, but computational overhead increases excessively
Solution Approach 1:
The patent segments the node update process into different stages based on node layers. Only first-layer nodes (direct neighbors of seed nodes) undergo full feature information updates in each iteration, while second-layer nodes and other nodes are processed with reduced updates or not updated at all. This segmented update strategy maintains classification accuracy for critical nodes while significantly reducing the total computational time required for processing the entire graph.
Solution Approach 2:
The patent applies partial action by performing complete feature updates only on the subset of nodes that are most critical for classification (seed nodes and first-layer nodes). Other nodes receive minimal or no updates, which is sufficient for maintaining overall classification accuracy while reducing the total computational time and overhead associated with processing all nodes in every iteration.
Data Source
AI summary
The present disclosure describes methods, apparatus, and storage medium for node classification and training a node classification model. The method includes obtaining a target node subset and a neighbor node subset corresponding to the target node subset from a sample node set labeled with a target node class, a neighbor node in the neighbor node subset being associated with a target node in the target node subset; extracting a feature subset of the target node subset based on the neighbor node subset by using a node classification model, the feature subset comprising a feature vector of the target node; performing class prediction for the target node subset according to the feature subset, to obtain a predicted class probability subset; and training the node classification model with a target model parameter according to the predicted class probability subset and a target node class subset of the target node subset.