Bipartite Interaction Graph Autoencoder for Unlabeled Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection models struggle to effectively model rich information in double-party interactions, as they either force data into tabular formats, losing relational properties, or fail to capture the rich attributes of customer-service provider interactions, and are unable to adapt to new fraud patterns without labeled data.
Innovation Solution
A graph convolutional neural network with an autoencoder structure processes bipartite interaction data, reconstructing node and edge attributes to derive anomaly scores, and employs a decoder to model relational data, allowing for the detection of anomalies based on reconstruction loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If existing anomaly detection models use tabular formats to represent interaction data, then the data processing is simplified, but the relational properties and rich attributes of interactions are lost
Solution Approach 1:
The patent transforms the data representation from traditional tabular format to graph structure, adding a dimensional aspect that preserves relational properties. The graph representation maintains both the tabular data processing capabilities and the relational structure through nodes, edges, and their attributes, effectively resolving the information loss problem while keeping processing feasible.
Solution Approach 2:
The patent creates a composite data structure that combines tabular attributes with graph relational structure. The hybrid representation integrates both the simplicity of tabular data and the richness of relational graphs, allowing the model to process data efficiently while preserving all interaction information including party attributes, interaction attributes, and relational patterns.
2Measurement precision
If existing models focus on labeled data for training, then detection accuracy for known patterns is improved, but the ability to adapt to new fraud patterns without labeled data deteriorates
Solution Approach 1:
The patent performs preliminary action by training the model on labeled historical fraud data to learn normal and anomalous interaction patterns. This preliminary training establishes a baseline detection capability that can then generalize to new, unseen fraud patterns through unsupervised anomaly detection, allowing the system to adapt without requiring labeled data for each new fraud type.
Solution Approach 2:
The patent implements feedback mechanisms where the model continuously learns from detected anomalies and adjusts its detection thresholds and patterns. The system uses feedback from both labeled training data and unlabeled operational data to refine its understanding of normal behavior, enabling it to adapt to new fraud patterns while maintaining accuracy for known patterns.
3Reliability
If complex graph structures with rich attributes are used to model interactions, then the ability to capture fraud patterns is improved, but the model complexity and computational requirements increase
Solution Approach 1:
The patent segments the complex fraud detection task into multiple processing stages: graph construction from interaction data, feature extraction from nodes and edges, anomaly scoring through comparison with normal patterns, and threshold-based detection. This segmentation breaks down the complex model into manageable components that can be processed efficiently while maintaining high detection capability.
Solution Approach 2:
The patent applies local quality by focusing computational resources on detecting anomalies at specific locations in the graph structure where fraud patterns are most likely to occur. The model prioritizes processing of interactions with higher anomaly scores and focuses on local graph neighborhoods, reducing overall computational requirements while maintaining detection reliability.
Data Source
AI summary
Aspects concern a method for detecting anomalies in double-party interaction data, comprising representing interactions between parties of a first group and parties of a second group as a graph, wherein each interaction between a first party of the first group and a second party of the second group is represented by an edge between a respective first node representing the first party and a respective second node representing the second party and wherein information about the first party is assigned to the first node as node attribute information, information about the second party is assigned to the second node as node attribute information and information about the interaction is assigned to the edge as edge attribute information, processing the graph by a graph convolutional neural network having an autoencoder structure, deriving anomaly scores for interactions, parties of the first group and parties of the second group from a reconstruction loss between the graph and an output of the graph convolutional neural network in response to the graph including at least a loss between the edge attribute information and edge attribute information reconstructed by a decoder of the graph convolutional neural network and detecting anomalies based on the anomaly scores.