Time-series Forecasting-based Financial Fraud Detection using Levy-Flight Distributed Dung Beetle Optimized Graph Convolutional LSTM Model

Abstract

This paper presents a novel Levy-Flight Distributed Dung Beetle Optimizer (LFDBO), designed to enhance financial fraud detection through time-series forecasting by addressing the limitations of traditional optimization methods. While previous models often struggle with non-convex optimization, our LFDBO introduces significant modifications to the existing Dung Beetle Optimizer (DBO) by incorporating Lévy flight random walks, which improve exploration efficiency within the hyperparameter space. This modified algorithm effectively balances exploration and exploitation, allowing for the avoidance of local optima and accelerating convergence speed. The LFDBO optimizes critical hyperparameters such as learning rate, batch size, and hidden units in Graph Convolutional Long Short-Term Memory (GC-LSTM) networks, resulting in superior fraud detection accuracy. Experimental results demonstrate that the LFDBO-optimized GC-LSTM model significantly outperforms traditional approaches, achieving a notable increase in detection precision across various synthetic financial datasets. This work not only contributes a novel algorithmic framework but also provides practical implications for enhancing financial security and decision-making.