Enhancing Financial Fraud Detection using LSTM- DNN with Hyena Optimization: A Comparative and Explainable Deep Learning Approach

Abstract

Financial Fraud Detection (FFD) have increase serious as fraudulent activities continue to change and pose significant risks to financial institutions and consumers alike. The increasing sophistication of fraudulent activities necessitates innovative approaches to identify and mitigate potential risks in financial transactions. The data preprocessing, containing data cleaning as well as handling of missing values, to get high-quality input for analysis. Feature engineering techniques are employed to create relevant attributes from raw data, improving the fraudulent patterns. A hybrid classification model utilizing Long Short-Term Memory and Deep Neural Network with Hyena Optimization Algorithm (LSTM-DNN with HOA) is implemented to seizure temporal dependencies in transaction data, improving detection accuracy. The dataset is split into training as well as testing sets to smooth verification of the model’s performance. Scaling normalization is applied to ensure all variables contribute equally to the analysis, whereas distance calculations help in assessing the similarity between transactions. After fraud prediction using the optimized LSTM-DNN model, Explainable Artificial Intelligence (XAI) is incorporated using Shapley Additive explanations (SHAP). SHAP is employed to take calculations by computing the support of each numerical feature toward fraud and non-fraud decisions, thereby improving transparency, trustworthiness, and decision interpretability. For FFD the Credit Card Fraud Detection dataset is implemented using python software the proposed LSTM-DNN with FOA have the accuracy of 99\% high compared to existing technique. Strength of the paper is the optimization process enhances model, convergence, and classification performance whereas reducing the chances of suboptimal parameter selection.