Deep Learning Paradigms for Multi-Dimensional Big Data Analytics: A Critical Assessment

Abstract

As the growth of big data has accelerated, there has been a need for advanced analytical methods to process this data and provide value from these assets. Deep learning algorithms have proven to be powerful in this context; they can model complex patterns. In this study, we assess the performance of convolutional neural networks (CNNs), recurrent neural networks (RNNs), and autoencoders in transaction fraud detection over the Kaggle “Credit Card Fraud Detection” dataset that contains more than 284000 transactions. The findings revealed that deep learning models were superior over traditional statistical approaches with respect to accuracies and scalability even though they need significant computational resources and careful tuning to handle class imbalance. This study adds to the burgeoning literature on big data analytics by exploring the advantages and disadvantages of deep learning methods for financial fraud detection in the world where access to big data has become ubiquitous for organizations. The Hybrid Model outperforms the rest with the highest estimated accuracy of 0.915, followed closely by the CNN and Autoencoder models, both around 0.85. The RNN (LSTM) also performs well with an accuracy of 0.82, while the baseline Logistic Regression model lags behind at 0.675.