On the Reusability of Machine Learning Models in Edge Computing: A Statistical Learning Approach

The adoption of Edge Computing continues to grow with edge nodes recording increasingly more data, which inevitably requires that they should be processed through Machine Learning (ML) models to speed up the production of knowledge. However, training these models requires an increased amount of resources, which are limited, thus, the reuse of ML models becomes of paramount importance. Given that we do not have a pool of models to choose from, is it possible to determine which nodes in the network require distinct models and which of them could be reused? In this paper, we propose a solution to this question, an online model reuse framework which is evaluated for its precision and speedup. The framework considers all possible combinations of pairs in the network to determine which are good reusability pairs, by adopting statistical learning methods. Then for each pair, the node model is chosen that has the highest data space overlap. Our comprehensive experimental analysis in the context of both regression and classification shows the feasibility our solution in model reusability in Edge Computing environments