Automatically Selecting Parameters for Graph-Based Clustering

Abstract

Data streams present a number of challenges, caused by change in stream concepts over time. In this thesis we present a novel method for detection of concept drift within data streams by analysing geometric features of the clustering algorithm, RepStream. Further, we present novel methods for automatically adjusting critical input parameters over time, and generating self-organising nearest-neighbour graphs, improving robustness and decreasing the need to domain-specific knowledge in the face of stream evolution.