Optimizing Fraud Detection Systems with Sequential Feature Analysis CNN and SVM Classification

Fraud detection in financial transactions has become one of the most challenging challenges because of the increased implementation of technological payment systems and the complexity of fraud schemes. The current work proposes a novel CNN-SVM model in handling the issues of identifying fraudulent activities in highly imbalanced datasets where the actual fraudulent transactions are as low as 0.3 % of the total. The hybrid method uses CNNs for capturing temporal and localized patterns from transaction sequences and the, patterns are accurately classified into fraudulent and non-fraudulent by using SVMs. The experimental evaluation has been conducted on a large transaction dataset with more than one million records. The hybrid model achieved 96.7%, outcompeting traditional models such as Logistic Regression (84.3<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>) and Random Forest (88.5<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>). It demonstrated higher recall rate (91.2%) and Fl-score (90.3%); it met the primary objective of lowering false negatives in fraud detection. The results further confirmed that the proposed hybrid model, CNNs improved the recall rate by 2.9 % and Fl-score by 2.6% in comparison to the individual CNNs, which indicates that the proposed model successfully the handle the unbalanced data. The hybrid model achieved computational efficiency suitable for practice since the time per inference was only 0.30 milliseconds. To this end, these investigation results demonstrate the utilization of the DL method in combination with the ML method, to address the limitations of each technique. By following the presented concept of utilizing a hybrid framework, today's financial systems can set a paradigm for scalable, precise, and efficient decision-making for fraud detection. Future expansions might examine how other non-supervised learning methods can also help decrease reliance on labeled images.