Analysis and visualization of fraud detection patterns through data mining and classification using MLP and hybrid deep learning model

Fraudulent transactions represent serious risks in various industries and need enhanced detection and classification methodologies. The investigation goes into the issue of unauthorized transaction classification, emphasizing the necessity of accurate identification to prevent monetary damage and preserve system integrity. The study employs powerful data mining and visualization approaches to discover hidden trends accompanying fraudulent behavior utilizing historical fraud data. This work develops two distinct deep learning models for classifying illicit transactions. The proposed model uses two reliable deep learning models, multilayer perceptron (MLP) and a hybrid model combining one-dimensional convolutional neural network (1D-CNN) and long short-term memory (LSTM), for classification. In addition, this work utilizes unique analytical visualizations and deep learning models to address the issue of bogus transaction classification. The considered dataset exhibits a built-in imbalance, with an uneven proportion of legal and wrongful transactions. The considered MLP performed significantly well, with an accuracy of 99.64%, a precision of 99.47%, a recall of 99.63%, and an F1-score of 99.64%. On the other hand, the proposed 1D-CNN-LSTM hybrid model has a more effective accuracy of 99.76%, precision of 98.94%, recall of 99.57%, and an F1-score of 99.46%. Therefore, the comparative evaluation of these models suggested their effectiveness in detecting fraudulent transactions, although a better knowledge of their performance is essential. While the MLP model is superior in precision and recall, the 1D-CNN-LSTM hybrid model improves accuracy and recall. The application’s particular requirements decide the selection between these classifications. If accuracy and recall are crucial, the MLP model might be employed, while the 1D-CNN-LSTM hybrid model arises as a promising rival if an appropriate ratio of accuracy and recall is required. The findings reveal the usefulness of the MLP and the 1D-CNN-LSTM frameworks, enabling a more thorough knowledge of their strengths and shortcomings. This allows sensible choices to be made depending on the application’s unique aims, enabling ongoing attempts to enhance systems for identifying fraudulent activity.