Sharing ML models using IoT communities of Interest based on data similarity and location

Abstract

The main goal of this master thesis is to develop an efficient method for sharing Machine Learning models across Internet of Things devices. To achieve this, the research work proposes a novel approach focused on distributing Machine Learning models among Internet of Things Communities of Interest based on the similarity of Internet of Things data streams and geospatial components such as location and elevation. To validate this approach, the study adopted a cluster-based strategy to form Internet of Things Communities of Interest. Initially, a thorough similarity analysis of IoT weather sensor data streams was conducted using both Dynamic Time Warping and Spearman’s correlation methods. Evaluation of the similarity results revealed that Spearman’s correlation performed better than Dynamic Time Warping, producing higher-quality and more coherent clusters. Thus, the study proceeded with K-means clustering using the outcomes of Spearman’s correlation analysis and goespatial data to form clusters, guided by the optimal number of clusters, four, determined through the elbow method. These clusters formed the foundation for Internet of Things - Communities of Interest, essential for the development, validation, testing, and sharing of Machine Learning models. Evaluation of Machine Learning model performance during the sharing and testing phases revealed that the majority of the Machine Learning models performed better when trained, tested, and shared within the same Community of Interest dataset. On the contrary, models trained on a different Community of Interest exhibited poorer performance when tested on members of another Community of Interest. The findings of this study demonstrate that it is possible to delineate geospatial zones based on the inherent similarity of Internet of Things data streams, and to craft and validate Machine Learning models tailored to the unique characteristics of each zone. It also establishes that it possible to leverage geospatial components for sharing and reusing pre-trained Machine Learning models among Internet of Things devices.