Analysis of web visit histories, part II: Predicting navigation by Nested STUMP Regression Trees

Abstract

This paper constitutes part II of the contribution to the analysis of web visit histories

through a new methodological framework for web usage-structure mining. The aim is to

detect which web patterns make a navigator survive on the web for more clicks with respect to another who leaves the web site earlier. The set of available web sections is fixed and can be related to one web portal, i.e. a dataset coming from the UCI Machine Learning Repository that can be considered as a real world case study. A tree-based approach is introduced to identify the prediction rule to explain the expected number of clicks in terms of web preferences click by click. As main result, we define an inedited regression tree procedure based on an hierarchy of STUMPs regression trees, where a STUMP is a tree with only one split or, equivalently, with only two terminal nodes. Suitable properties of the proposed method are outlined. This novel tree-structure can be considered as an alternative to the standard regression tree and can be applied in various contexts different from web mining.