Exploratory Cluster Analysis from Ubiquitous Data Streams using Self-Organizing Maps

Abstract

This thesis addresses the use of Self-Organizing Maps (SOM) for exploratory cluster analysis over ubiquitous data streams, where two complementary problems arise: first, to generate (local) SOM models over potentially unbounded multi-dimensional non-stationary data streams; second, to extrapolate these capabilities to ubiquitous environments. Towards this problematic, original contributions are made in terms of algorithms and methodologies. Two different methods are proposed regarding the first problem. By focusing on visual knowledge discovery, these methods fill an existing gap in the panorama of current methods for cluster analysis over data streams. Moreover, the original SOM capabilities in performing both clustering of observations and features are transposed to data streams, characterizing these contributions as versatile compared to existing methods, which target an individual clustering problem. Also, additional methodologies that tackle the ubiquitous aspect of data streams are proposed in respect to the second problem, allowing distributed and collaborative learning strategies. Experimental evaluations attest the effectiveness of the proposed methods and realworld applications are exemplified, namely regarding electric consumption data, air quality monitoring networks and financial data, motivating their practical use. This research study is the first to clearly address the use of the SOM towards ubiquitous data streams and opens several other research opportunities in the future.