Abstract:
Frequent graph mining is a challenging task that extracts novel and useful knowledge from graph data. The problem becomes even more challenging when the information comes from data streams which evolve in real-time. Additionally, concept drift detection is needed for the problem domain. In this thesis we present an approach for adaptively mining frequent graph patterns on time-varying streams. Our approach extends an existing graph batch mining framework by implementing and integrating a state-of-the-art change detector. The approach works on coresets of closed frequent subgraphs, compressed representations of graph sets and uses the change detector to address potential concept drifts. In our approach, we mine and monitor the concept drifts of the coresets of closed frequent subgraphs. An evaluation study on large scale datasets compares the performance between our approach and a change-adaptive algorithm in the existing graph batch mining framework. The experiments process different real-world chemical molecular and social network graph datasets with varying severity of artificial drifts.