Graph Feature Calculation via Node Segmentation and Interpolation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for calculating feature values from graphs representing communication networks face increased temporal and spatial calculation loads as the number of nodes grows.
Innovation Solution
A feature value calculation device that generates a graph representing inter-node communication, selects nodes based on predetermined conditions, calculates feature values using a specific learning method for selected nodes, and estimates feature values for other nodes by combining values from sequentially adjacent nodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional graph embedding methods (DeepWalk, Node2Vec) are used to calculate feature values for all nodes, then the quality of feature values improves, but temporal and spatial calculation loads increase significantly
Solution Approach 1:
The patent segments the graph nodes into two categories: representative nodes (seed nodes) and non-representative nodes. Feature values are calculated exactly only for representative nodes using graph embedding methods, while feature values for non-representative nodes are estimated through interpolation or extrapolation based on their representative nodes. This segmentation allows high-quality feature calculation to be applied selectively rather than universally, reducing overall calculation time while maintaining feature quality for the most important nodes.
Solution Approach 2:
The patent extracts and identifies a subset of representative nodes from the complete graph that capture the essential structural characteristics. By focusing computational resources on calculating feature values only for these extracted representative nodes rather than all nodes, the method achieves significant time savings while preserving the ability to derive meaningful features for the entire graph through the representative nodes.
2Measurement precision
If conventional graph embedding methods are used to calculate feature values for all nodes, then comprehensive feature coverage improves, but spatial calculation loads increase
Solution Approach 1:
The patent segments nodes into representative and non-representative categories, storing computed feature values only for representative nodes in memory. Non-representative node features are computed on-demand through interpolation from representative nodes, reducing the static memory footprint while maintaining comprehensive feature coverage when needed.
3Measurement precision
If feature values are calculated for all nodes using learning methods, then accuracy for all nodes improves, but calculation complexity increases
Solution Approach 1:
The patent applies different calculation strategies to different node segments: representative nodes undergo full learning-based feature calculation, while non-representative nodes use simpler interpolation or extrapolation methods based on their representative nodes. This differential approach reduces overall calculation complexity while maintaining accuracy where it matters most.
Solution Approach 2:
The patent creates feature estimates for non-representative nodes by copying and combining feature information from representative nodes through interpolation or extrapolation, rather than performing full learning calculations for each node. This copying approach significantly reduces calculation complexity while providing reasonable feature estimates for all nodes.
Data Source
AI summary
A generation unit is configured to generate a graph representing inter-node communication using information on communication between nodes on a network. A selection unit is configured to select a node satisfying a predetermined condition among nodes in the generated graph. A calculation unit is configured to calculate a feature value in the graph for the selected node by a predetermined learning method. An estimation unit is configured to estimate a feature value for a node other than the selected nodes by combining feature values estimated for sequentially adjacent nodes in the graph.