Adapt but Do Not Forget: Towards Enhancing Drift Handling in 6G Networks

The advent of 6G represents a paradigm shift where Artificial Intelligence and Machine Learning AI/ML becoming the pulsating heart of network innovation. However, these AI/ML models face a significant challenge known as Model Drift, caused by changes in data distributions and feature relationships over time, leading to degraded performance. While numerous drift adaptation strategies have been explored including model retraining, incremental learning, and ensemble approaches, they often neglect the critical issue of "Catastrophic Forgetting" (CF). The latter refers to the phenomenon where neural networks, upon learning new information (post-drift), lose the ability to retain previously learned knowledge (pre-drift). This challenge is particularly pronounced in the time-series forecasting tasks inherent to telecommunications networks, where sequential data dependencies amplify the forgetting issue. To address this limitation, we propose a novel LSTM-EWC model that synergistically combines the time-series processing strengths of Long Short-Term Memory (LSTM) networks with the CF mitigation capabilities of Elastic Weight Consolidation (EWC). The proposed framework empowers LSTM models to retain essential pre-drift knowledge, while effectively adapting to post-drift changes, enabling robust and continuous learning in dynamic environments. Experimental evaluations on a real-world telecommunications dataset highlight the efficacy of our approach, demonstrating up to 51% improvement in knowledge retention over state-of-the-art CF baseline methods.