Representing models for local sequential behaviors in temporal sequence database

Temporal sequences that are lists of transaction records ordered by transaction time constitute a large part of data stored in many information systems. Recently, there has been a lot of interest in using data mining techniques to extract frequent patterns from temporal sequences in various applications. Previous work mainly considered global sequential behaviors. The local patterns, of which the frequency is large only in a subsequence of the original sequence, are very common in practice, whereas have not received enough attention in KDD field. Therefore the formal definition of the representing model for local sequential behaviors is of crucial importance, since it can help further designing efficient mining algorithm for this knowledge. In this paper, we propose four representing models for local sequential behaviors with various mining goals, which include single-record patterns, multi-record patterns, non-trivial patterns, and generalized patterns. In addition, to facilitate the discovery of this knowledge, we also propose a simplified model, which enables the implementation of the method that scales linearly. Some experimental examples are given, which clarify the definition of our representing models and demonstrate kinds of novel knowledge.