DeepFM-MOTSO: A Deep Factorization Machine Framework Optimized by Multi-Objective Tuna Swarm for Online Advertising Fraud Detection

Abstract

Online advertising fraud is still a significant problem that inflates marketing expenses and distorts campaign performance. In order to improve fraud detection accuracy in the face of extreme class imbalance, this study proposes a hybrid DeepFM–MOTSO framework that combines a Deep Factorization Machine (DeepFM) with a Multi-Objective Tuna Swarm Optimizer (MOTSO). While MOTSO simultaneously optimizes several goals—maximizing F1-score, precision, recall, and AUC, as well as minimizing loss—for balanced classification, the model captures both low-order feature interactions through the FM component and high-order nonlinear representations via deep neural layers. A real-world advertising dataset with roughly 2,043 records that included contextual and behavioral features like session_duration, click_interval, impression_count, and device_type was used to test the method. With an accuracy of 0.952, precision of 0.214, recall of 0.179, F1-score of 0.195, and AUC of 0.864, the experimental results demonstrate that the proposed DeepFM–MOTSO outperformed all comparative baselines, indicating superior capability in identifying minority-class fraudulent instances. The findings confirm that multi-objective optimization effectively improves model convergence, stability, and real-time adaptability for intelligent online advertising fraud detection.