Accelerated Artificial Bee Colony Optimization for Cost-Sensitive Neural Networks in Multi-Class Problems

Hacilar H., DEDETÜRK B. K., ÖZMEN M., Celik M. E., Gungor V. C.

Expert Systems, vol.42, no.5, 2025 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 42 Issue: 5
  • Publication Date: 2025
  • Doi Number: 10.1111/exsy.70045
  • Journal Name: Expert Systems
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Keywords: anomaly detection, artificial bee colony, artificial neural network, cost-sensitive learning, earthquake damage estimation, GPU parallelization, network intrusion detection, swarm intelligence, UNSW-NB15
  • Erciyes University Affiliated: Yes

Abstract

Metaheuristics are advanced problem-solving techniques that develop efficient algorithms to address complex challenges, while neural networks are algorithms inspired by the structure and function of the human brain. Combining these approaches enables the resolution of complex optimization problems that traditional methods struggle to solve. This study presents a novel approach integrating the ABC algorithm with ANNs for weight optimization. The method is further enhanced by vectorization and parallelization techniques on both CPU and GPU to improve computational efficiency. Additionally, this study introduces a cost-sensitive fitness function tailored for multi-class classification to optimize results by considering relationships between target class levels. It validates these advancements in two critical applications: network intrusion detection and earthquake damage estimation. Notably, this study makes a significant contribution to earthquake damage assessment by leveraging machine learning algorithms and metaheuristics to enhance predictive models and decision-making in disaster response. By addressing the dynamic nature of earthquake damage, this research fills a critical gap in existing models and broadens the understanding of how machine learning and metaheuristics can improve disaster response strategies. In both domains, the ABC-ANN implementation yields promising results, particularly in earthquake damage estimation, where the cost-sensitive approach demonstrates satisfactory outcomes in macro-F1 and accuracy. The best results for macro-F1, weighted-F1, and overall accuracy provides best results with the UNSW-NB15 and earthquake datasets, showing values of 64%, 72%, 68%, and 60%, 80%, and 79%, respectively. Comparative performance evaluations reveal that the proposed parallel ABC-ANN model, incorporating the novel cost-sensitive fitness function and enhanced by vectorization and parallelization techniques, significantly reduces training time and outperforms state-of-the-art methods in terms of macro-F1 and accuracy in both network intrusion detection and earthquake damage estimation.