Efficient Approximate Mining of Frequent Patterns over Transactional Data Streams

Abstract

We investigate the problem of finding frequent patterns in a continuous stream of transactions. It is recognized that the approximate solutions are usually sufficient and many existing literature explicitly trade off accuracy for speed where the quality of the final approximate counts are governed by an error parameter, ε. However, the quantification of ε is never simple. By setting a small ε, we achieve good accuracy but suffer in terms of efficiency. A bigger ε improves the efficiency but seriously degrades the mining accuracy. To alleviate this problem, we offer an alternative which allows user to customize a set of error bounds based on his requirement. Our experimental studies show that the proposed algorithm has high precision, requires less memory and consumes less CPU time.