The main challenge of mining sequential patterns is the high processing cost of support counting for large amount of candidate patterns. For solving this problem, SPAM algorithm was proposed in SIGKDD'2002, which utilized a depth-first traversal on the search space combined with a vertical bitmap representation to provide efficient support counting. According to its experimental results, SPAM outperformed the previous works SPADE and PrefixSpan algorithms on large datasets. However, the SPAM algorithm is efficient under the assumption that a huge amount of main memory is available such that its practicability is in question. In this paper, an Improved-version of SPAM algorithm, called I-SPAM, is proposed. By extending the structures of data representation, several heuristic mechanisms are proposed to speed up the efficiency of support counting further. Moreover, the required memory size for storing temporal data during mining process of our method is less than the one needed by SPAM. The experimental results show that I-SPAM can achieve the same magnitude efficiency and even better than SPAM on execution time under about half the maximum memory requirement of SPAM.
展开▼
