Graph Neural Network Embedding for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There has been limited success in adapting deep anomaly detection methods to graph-level anomaly detection, as existing approaches often perform close to random and fail to effectively detect entire abnormal graphs, rather than localized anomalies within graphs.
Innovation Solution
The One-Class Graph Transformation Learning (OCGTL) approach combines deep one-class classification and self-supervision using K+1 graph neural networks, jointly trained on complementary deep anomaly detection losses to overcome issues of mode collapse and hypersphere collapse, enabling the detection of abnormal graph embeddings through both Euclidean and angular distances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing deep anomaly detection methods are applied to graph-level data, then the detection process can be performed, but the detection accuracy deteriorates to near random performance
Solution Approach 1:
The patent transforms the anomaly detection problem by changing the parameter space from direct graph comparison to embedding space representation. Graphs are transformed into vector embeddings through neural networks, allowing anomaly detection to operate in a continuous parameter space where distance metrics (Euclidean and angular) can effectively measure deviations from normal patterns, thereby resolving the near-random performance issue of direct graph-level detection
Solution Approach 2:
The patent introduces graph embeddings as an intermediary representation between raw graph data and anomaly detection. Instead of directly analyzing graph structures, the system uses embedding vectors as a mediator that captures essential graph characteristics in a form suitable for anomaly scoring, enabling effective detection without requiring complex graph-level operations
2Reliability
If multiple graph neural networks are used to improve detection robustness, then the detection capability improves, but the computational complexity increases
Solution Approach 1:
The patent segments the anomaly detection task into multiple independent graph neural networks, each responsible for producing embeddings from different perspectives or feature extractors. This segmentation allows each network to be relatively simple while the ensemble of segmented networks provides robust detection, balancing complexity and performance by dividing the overall system into manageable components
Solution Approach 2:
The patent designs graph neural networks with universal feature extraction capabilities that can handle various graph types and anomaly patterns through a common embedding framework. The same network architecture serves multiple functions: feature extraction, embedding generation, and anomaly scoring, reducing overall system complexity while maintaining robust detection across different scenarios
Data Source
AI summary
A computer-implemented method of anomaly detection associated with graphical data includes receiving as input one or more input data sets, wherein the input data sets includes one or more graphs, utilizing a plurality of graph neural networks (GNNs) to identify an aggregate loss including a first loss and second loss associated with the input data set, wherein the aggregate loss is associated with embedding's of the GNNs, and outputting a notification associated with an anomaly detection when the first and second loss exceeds an anomaly-detection threshold.