ML Alert Review Using Graph Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current alert review processes in regulatory settings, such as Anti-Money Laundering (AML), are cumbersome and inefficient due to the complexity of navigating large networks of financial transactions, high false positive rates, and the lack of context-aware tools for analysts, leading to challenges in identifying suspicious activities.
Innovation Solution
A machine learning-based alert review system using self-supervised graph representation learning and graph neural networks to model transactions and customers as a heterogeneous graph, providing context-aware representations and interactive visual interfaces for analysts to cluster and analyze transactions, reducing the burden of manual data aggregation and improving detection of suspicious activities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning is used to detect security issues and generate alerts, then detection accuracy is improved, but the number of alerts increases leading to cumbersome manual review
Solution Approach 1:
The patent segments the alert review process by introducing multiple analyst roles (first analyst and second analyst) who independently review different aspects of alerts. This division allows parallel processing of alert evaluation, reducing the time burden on individual analysts while maintaining detection accuracy through collaborative review.
Solution Approach 2:
The system performs preliminary actions by automatically generating alerts with pre-computed features and embeddings before human review. The graph neural network pre-processes transaction data to create structured representations, enabling analysts to focus their time on evaluating pre-processed alert information rather than raw data aggregation.
2Loss of information
If analysts manually review large networks of transactions, then context awareness is improved, but the complexity and time required for review increases
Solution Approach 1:
The patent introduces graph neural networks as an intermediary between raw transaction data and analyst review. The GNN automatically constructs heterogeneous graphs that capture contextual relationships among transactions, entities, and accounts, providing analysts with pre-computed contextual insights without requiring them to manually navigate complex transaction networks.
Solution Approach 2:
The system replaces the mechanical process of manual data aggregation and context analysis with automated graph-based machine learning. The graph neural network automatically computes contextual representations and generates alert features, substituting the manual mechanical review process with intelligent automated processing that reduces analyst burden while improving context awareness.
3Reliability
If traditional rule-based systems are used for alert generation, then false positive rates are high, but system complexity is low
Solution Approach 1:
The patent changes the parameters of alert generation by transitioning from simple rule-based thresholds to graph neural network-based feature extraction. The system computes contextual embeddings and graph-based features that capture complex relationships among transactions, enabling more accurate alert scoring and reduced false positives through enriched parameter representation rather than basic rule matching.
Solution Approach 2:
The system employs a composite approach by combining graph neural networks with traditional alerting mechanisms. The heterogeneous graph structure integrates multiple data types (transactions, entities, accounts) and the GNN output is combined with rule-based filtering, creating a composite system that leverages both automated contextual analysis and established regulatory rules to reduce false positives while managing complexity.
4Quantity of substance
If more transactions and entities are associated with each alert, then detection completeness is improved, but the burden on analysts increases
Solution Approach 1:
The system performs preliminary aggregation and structuring of large volumes of transaction data into graph-based representations before presenting alerts to analysts. The graph neural network pre-computes contextual relationships and summarizes transaction networks, enabling analysts to review consolidated alert information that incorporates comprehensive transaction volumes without manually processing each individual transaction.
Solution Approach 2:
Instead of presenting raw transaction volumes to analysts and expecting them to find patterns, the system inverts the approach by using graph neural networks to automatically discover patterns and relationships within large transaction volumes, then presenting synthesized alert information to analysts. This reversal reduces analyst burden by having the system perform the complex analysis of large data volumes first.
Data Source
AI summary
In various embodiments, a process for alert review using machine learning and interactive visualizations includes receiving transaction data for transactions, using a machine learning model to determine embedding representations of the transaction data, and using one or more automated rules to identify of a subset of the transactions. The process includes using at least a portion of the embedding representations to automatically cluster the identified subset of the transactions into a plurality of different cluster groups, and providing an interactive visual representation of the plurality of different cluster groups.