Mining Emerging Patterns of Data Stream for Improving Classification Performance

Emerging patterns consider the support of patterns in the target class and opposite class datasets, which has a significant impact on improving the performance of data stream classification. The emerging pattern must first be a frequent pattern, and IncMine is a well-known data stream algorithm with frequent closed pattern mining. However, existing frequent pattern mining algorithms are oriented toward transactional data streams, and the itemset does not have class constraints, making it impossible to further mine emerging patterns for data stream classification. This paper proposes an emerging patterns based on IncMine (EPBIM) mining method, which improves the IncMine algorithm and uses Charm_Cla to mine frequent closed itemsets with class value constraints and obtain emerging patterns for Bayesian classification of data streams. The Charm_Cla runs on the WEKA platform, mining the emerging patterns of a batch and updating the semi-emerging patterns (semi-EPs) in the sliding window of data stream through the IncMine algorithm. Multiple real and simulated data streams were run on the data stream analysis platform massive online analysis (MOA), and the experimental results verified the effectiveness of the method.