Graph Neural Network Transaction Classification Framework
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computer models struggle to accurately detect fraudulent transactions due to dynamic fraudulent tactics that involve slight variations in transaction attributes, leading to confusion and inaccurate pattern derivation.
Innovation Solution
A machine learning model framework that utilizes multiple graph analysis techniques to analyze both actual and fuzzy attributes of transactions, along with a community aspect, to classify data more accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional computer models classify transactions based on identical attributes, then detection accuracy is improved for known fraudulent patterns, but detection accuracy deteriorates when fraudulent tactics use slightly different attributes
Solution Approach 1:
The patent transforms discrete transaction attributes into continuous embedding vectors through graph neural networks. This parameter transformation allows the system to capture semantic similarities between different attribute values (e.g., different IP addresses from the same range) and dynamically adapt to fraudulent tactics by learning relationships in the continuous vector space rather than relying on exact attribute matches
Solution Approach 2:
The patent adds a new dimension by introducing graph-based embedding vectors alongside traditional transaction attributes. This multi-dimensional approach combines structured attribute data with unstructured relationship data from transaction graphs, enabling the model to detect fraud based on both attribute values and their contextual relationships, thus improving adaptability to evolving fraudulent patterns
2Device complexity
If computer models rely on exact attribute matching, then classification simplicity is maintained, but pattern recognition capability deteriorates when attributes vary slightly
Solution Approach 1:
The patent introduces graph embedding vectors as an intermediary between raw transaction attributes and the classification model. These embeddings serve as a bridge that translates varied attribute values into a unified representation space, allowing the model to recognize patterns based on semantic similarity rather than exact matches, thus improving reliability without significantly increasing complexity
Solution Approach 2:
The patent replaces the mechanical exact-matching system with a learning-based embedding system. Instead of using rigid attribute equality checks, the system uses neural network-based embeddings that automatically learn meaningful representations and similarities, substituting a flexible adaptive mechanism for a rigid deterministic one
Data Source
AI summary
Methods and systems are presented for providing a machine learning model framework that uses multiple models that analyze different aspects of graph data to perform transaction classification. A graph is generated to represent relationships among transactions and fuzzy attributes. The framework includes a graph neural network that generates embeddings for each transaction based on the graph. The framework further includes a machine learning model that generates an initial classification score for a particular transaction based on the embeddings generated for the particular transaction and the actual attributes associated with the particular transaction. One or more communities are identified within the graph based on the connections among various fuzzy attributes. Characteristics associated with a particular community corresponding to the particular transaction are used to modify the initial risk score. A classification is determined for the particular transaction based on the modified risk score.