An empirical evaluation of classifier combination schemes for predicting user navigational behavior

Abstract

In this paper we present a simple classification system for predicting user behavior when browsing a Web site devoted to informing about university degrees. More than building a very accurate classifier, we want to study which kind of combination scheme performs better in front of a complexity constraint. A set of marks embedded in the Web pages being visited by each user is used as the input for a classification system which decides whether the user will be interested in accessing other related parts of the Web site or not. We compare two different classification systems: the first one is built using decision trees for the whole data set, with the aim of studying user profiles and variable importance, while the second one combines simple classifiers based on small decision trees using a combination of the voting and cascading paradigms, in order to make predictions which evolve during the period of time the Web site is collecting data. Results show that it is possible to extract useful information for studying user profiles and for predicting user behavior using small decision trees.