Effective AI/ML Training using Submodular Cell Selection for Network Energy Saving

Data collection from multiple cells is essential to build AI/ML based models in next generation 5G/6G cellular networks. However, pulling data from high percentage of the cells in the network to train a model, would require high training complexity overhead comprising of network (or spectrum) bandwidth, compute and storage. From the network field data, we observe the data distribution from multiple cells are not independent and identically distributed (i.i.d.), but are rather biased in the region of operations, and using data from high percentage of cells in the network to train the AI/ML models, as done in some previous works, may be suboptimal. Therefore, an efficient cell selection scheme for data collection to build AI/ML is critical for wireless cellular systems to optimize model performance while requiring low training complexity overhead. In this work, we propose novel schemes which incorporate submodular functions structure to select the cells for gathering training data. For the network energy saving use case, we train load forecasting model on cells selected from our proposed approach, and demonstrate upto 2X advantage in terms of top 5 % cell throughput drop performance while having similar power saving w.r.t. the baseline scheme of collecting training data from all the cells in the network. Additionally, we show up to 90X reduction in training data collection using our proposed scheme.