Abstract:
Online mining is a difficult task especially when such data streams evolve over time. Evolving data stream occurs when concepts drift or change completely, is becoming one of the core issues. A large portion of change detection research are carried out in the area of supervised learning, very little has been carried out for unlabeled data specifically in the area of transactional data streams. Overall when we monitor changes in transactional data we can consider two different types of changes: local and global change. Local changes are changes in distribution of the data, whereas global changes are data composition changes within the data stream. To detect changes in transactional data streams containing unlabeled data, we introduce a new technique called CD-TDS, that detects both these changes. Our change detector can identifies changes in relationships between items as data evolves with the progression of a stream. Crucially, detection of global drift enables us to better understand the dynamics in relationships that takes place over time. Experimental results using both real world and synthetic data show that the proposed approach is robust to noise and identifies structural changes with a high true positive rate while preserving a low false alarm rate.