Optimizing LGBM for Multi-Classification of 5G SA Traffic

In recent years, with the widespread adoption of 5G Standalone (SA) technology in mobile networks, there has been an increasing need to address the complex and data-intensive nature of 5G SA traffic classification. Building upon previous research, this paper enhances the performance of the Light Gradient Boosting Machine (LGBM) for classifying 5G SA traffic, using physical channel records as input. This approach helps reduce dataset dimension and addresses user privacy concerns. The paper also focuses on mitigating overfitting - ensuring the model's generalization ability - and incorporates zero-shot transfer learning techniques. We have refined LGBM by integrating dropout, regularization, and specialized feature engineering, which significantly boosts the model's performance on unseen data. Validated on a comprehensive dataset designed to reflect real-world 5G traffic scenarios, our optimized model achieves an overall accuracy of 89% in multi-class classification on unseen data across four classes and eleven different case scenarios, markedly improving upon the baseline accuracy of 66% observed with other methods.