Data Mining based Hybrid Latent Representation Induced Ensemble Model Towards Fraud Prediction

The increasing growth of e-commerce and mobile payments has contributed to a rise in the effects of financial payment services fraud over the last few years. "At 2025, the losses in fraud around the world could amount to about $44 billion, according to McKinsey." Every year, trillions of dollars are lost from fraudulent card transactions. An efficient fraud detection system based on advanced machine learning techniques have gained excellent value in all Financial Institutions to reduce these losses. The proposed algorithm SLREnsemble (Stochastic Latent Representation Ensemble) is a hybrid model that uses deep neural network Autoencoders to obtain the latent representation of genuine and fraudulent transactions and these representations are given as an input to the balanced ensemble model. The technique is tested against other existing methods, i.e., Isolation Forest, Random Forest, and XGBoost algorithms, by various performance metrics such as accuracy, precision, and recall. The results prove that the SLREnsemble model is highly efficient in Fraud detection and performs well with minimal training data and yields good precision, recall and f1-score with high computational speed values.