FINGERPRINT

Abstract

Monitoring and interpretation of changing patterns is a task of paramount importance for data mining applications in dynamic environments. While there is much research in adapting patterns in the presence of drift or shift, there is less research on how to maintain an overview of pattern changes over time. A major challenge is summarizing changes in an effective way, so that the nature of change can be understood by the user, while the demand on resources remains low. To this end, the authors propose FINGERPRINT, an environment for the summarization of cluster evolution. Cluster changes are captured into an “evolution graph,” which is then summarized based on cluster similarity into a fingerprint of evolution by merging similar clusters. The authors propose a batch summarization method that traverses and summarizes the Evolution Graph as a whole and an incremental method that is applied during the process of cluster transition discovery. They present experiments on different data streams and discuss the space reduction and information preservation achieved by the two methods.