XAI-driven Model Design for Resource Utilization Forecasting in Cloud-native 6G Networks

As cloud-native 6th Generation (6G) networks emerge, the resource utilization forecasting becomes crucial for effective service and network orchestration. While Artificial Intelligence (AI) holds promise in this domain, the diverse nature of the 6G underlying infrastructure and services poses significant challenges on the customization and the efficient design of the AI models. In this paper, we introduce the adoption of eXplainable AI (XAI) to generate spatio-temporal insights on the predictions of advanced AI models. Additionally, we present DuCAT, a Dual Cumulative Attribution Thresholding (DuCAT) heuristic algorithm, for feature and time window size selection towards AI model reduction. Experimental results on a publicly available dataset of cloud resource traces demonstrate that our proposed approach can efficiently reduce the AI model’s complexity (up to 60% decrease in inference time) without compromising prediction accuracy, addressing critical requirements for agile and resource-efficient 6G networks.