Structural Graph Neural Networks for Dynamic Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anomaly detection in dynamic graphs is challenging due to the complexity of data and variations in graph structure, as anomalous edges cannot be determined from a single timestamp and vertical sets change over time, requiring consideration of previous graphs for detection.
Innovation Solution
A structural graph neural network framework is proposed to detect anomalous edges by learning graph structure changes within a given time window, using subgraph structure generation, graph structure feature extraction, and a detection network to predict edge categories, incorporating recurrent neural networks for temporal information capture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning methods are used for anomaly detection in dynamic graphs, then the detection can be performed with simpler algorithms, but the detection accuracy is insufficient due to the complexity of data and variations in graph structure
Solution Approach 1:
The patent segments the dynamic graph analysis into multiple time windows, where each window captures graph structures at different timestamps. This segmentation allows the model to analyze temporal variations systematically, improving detection accuracy by comparing structures across windows while managing complexity through localized analysis rather than processing the entire dynamic graph at once.
Solution Approach 2:
The patent transforms the graph structure analysis by adding a temporal dimension through time windows. Instead of analyzing a single static graph, the method creates multiple graph representations across different time points, enabling the detection of structural changes and anomalies that vary over time. This dimensional transformation converts a complex temporal problem into a series of comparable spatial structures.
2Measurement precision
If graphs from multiple timestamps are considered for anomaly detection, then the detection accuracy improves, but the computational cost increases
Solution Approach 1:
The patent applies partial action by focusing computational resources on specific time windows and relevant graph substructures rather than processing all possible graph combinations. The method selectively analyzes graphs within defined time windows and uses graph neural networks to process only the necessary structural information, reducing overall computational cost while maintaining detection accuracy through targeted analysis.
Solution Approach 2:
The patent performs preliminary action by pre-processing graph structures into standardized representations before anomaly detection. Graph neural networks are used to extract meaningful features and representations in advance, transforming raw graph data into a format that is more efficient for subsequent anomaly detection. This preliminary feature extraction reduces the computational burden during the actual detection phase.
3Measurement precision
If vertical sets are updated along all timestamps, then the detection remains current and accurate, but the processing time increases
Solution Approach 1:
The patent implements periodic action by analyzing graph structures at discrete time windows rather than continuously updating at every timestamp. This periodic sampling approach captures essential temporal variations while reducing processing time by skipping intermediate updates. The method balances detection accuracy with processing efficiency by selecting strategic time points for analysis.
Solution Approach 2:
The patent uses preliminary action through pre-computed graph representations and cached structural features from previous time windows. Instead of re-processing all graph data at each timestamp, the method leverages previously extracted features and updates only the necessary components, significantly reducing processing time while maintaining detection accuracy through incremental updates.
Data Source
AI summary
A computer-implemented method for graph structure based anomaly detection on a dynamic graph is provided. The method includes detecting anomalous edges in the dynamic graph by learning graph structure changes in the dynamic graph with respect to target edges to be evaluated in a given time window repeatedly applied to the dynamic graph. The target edges correspond to particular different timestamps. The method further includes predicting a category of each of the target edges as being one of anomalous and non-anomalous based on the graph structure changes. The method also includes controlling a hardware based device to avoid an impending failure responsive to the category of at least one of the target edges.