Strategic Sampling: A Framework for Enhancing Speed and Performance of Financial Fraud Detection Models

Abstract

Financial fraud detection is a high-stakes field where rapid inference is essential. While state-of-the-art fraud detection models vary in terms of architectural decisions and appear to exhibit unique computational bottlenecks, we highlight that their run-times are all dominated by extensive information-gathering steps. These steps involve aggregating information from a large set of nodes or edges within a graph, and these intensive steps are performed O(|V |) or O(|E|) times during an inference forward pass, on a graph with |V | nodes and |E| edges. We introduce Strategic Sampling, a general method to accelerate these information-gathering steps. Our approach tailors sampling strategies based on the specific objective function used in each model’s information-gathering process, selecting the most relevant pieces of information to use in each step. This ensures that critical information is retained while significantly reducing the amount of data processed, thus speeding up the computation. We conceptually demonstrate how Strategic Sampling can be applied to message-passing Graph Neural Networks, Graph Transformers, and TGEditor (a state-of-the-art graph editing algorithm). To showcase the effectiveness of our proposed Strategic Sampling method, we implement it in the TGEditor codebase. Our results show that Strategic Sampling not only significantly reduces computation time by more than an order of magnitude, but also improves the F1 score, enhancing both efficiency and performance. This study underscores the potential of Strategic Sampling to universally boost the performance of various financial fraud detection models, paving the way for faster and more accurate fraud detection.