Document Type


Publication Date



Mining frequent patterns is an important component of many prediction systems. One common usage in web applications is the mining of users' access behavior for the purpose of predicting and hence pre-fetching appropriate web pages.

Mining solutions in the literature are often based on the use of an Apriori-like candidate generation strategy, which typically requires numerous scans of the potentially huge sequence database. In this paper we instead introduce a more efficient strategy for discovering frequent patterns in sequence databases that requires only two scans of the database. The first scan obtains support counts for subsequences of length two. The second scan extracts potentially frequent sequences of any length and represents them as a compressed patter tree structure (FS-tree). Frequent sequence patterns are then mined from the FS-tree. Incremental and interactive mining functionalities are also facilitated by the FS-tree. Our FS-Miner system has the ability to adapt to changes in users' behavior over time, in the form of new input sequences, and to respond incrementally without the need to perform full re-computation. Our system also allows the user to change the input parameters (e.g., minimum support and desired pattern size) interactively without requring full re-computation in most cases.

We have tested out system using two different data sets, comparing it against two other algorithms from the literature. Our experimental results show that our system scales up linearly with the size of the input database. Furthermore, it exhibits excellent scalability with respect to support threshold decreases. We also show that the incremental update capability of the system provides significant performance advantages over full re-computation even for relatively large update sizes.