An Empirical Study of Extracting Multidimensional Sequential Rules for Personalization and Recommendation in Online Commerce

The application of web mining to personalization has a long tradition in electronic commerce research. In this empirical study we focus specifically on mining sequential navigation patterns from weblogs and thoroughly compare different design variants for making personalized suggestions to users. In particular we concentrate on the impact of additional product knowledge like item characteristics, different properties of the sequential pattern mining process such as closure as well as rule quality metrics such as support, confidence and lift, and evaluate the recommender’s accuracy by experimenting on historical web sessions. This paper therefore firstly demonstrates how state of the art sequence mining algorithms such as PrefixSpan and BIDE may be adapted to the specific problem of extracting sequential rules from e-commerce weblogs. Furthermore, in order to compact the resulting rule set, the ∆-closed criteria is proposed as a logical extention to closed and maximal frequent patterns to eliminate spurious sequences. Finally, our experimental findings show that using multidimensional sequential patterns and the lift metric for weighting personalization rules can boost recall to 28% of all actual purchase transactions when using only short navigational sequences.