Investigation of sequential pattern mining techniques for web recommendation

Increased application of sequence mining in web recommender systems (WRS) requires a better understanding of the performance and a clear identification of the strengths and weaknesses of existing algorithms. Among the commonly used sequence mining methods, the tree-based approach, such as pre-order linked WAP-tree mining algorithm (PLWAP-Mine) and conditional sequence mining algorithm (CS-Mine), has demonstrated high performance in web mining applications. However, its advantages over other mining methods are not well explained and understood in the context of WRS. This paper firstly reviews the existing sequence mining algorithms, and then studies the performance of two outstanding algorithms, i.e., the PLWAP-Mine and CS-Mine algorithms, with respect to their sensitivity to the dataset variability, and their practicality for web recommendation. The results show that CS-Mine performs faster than PLWAP-Mine, but the frequent patterns generated by PLWAP-Mine are more effective than CS-Mine when applied in web recommendations. These results are useful to WRS developers for the selection of appropriate sequence mining algorithms.