On the right track! Analysing and predicting navigation success in Wikipedia

Abstract

Understanding and modeling user navigation behaviour in the web is of interest for different applications. For example, e-commerce portals can be adjusted to strengthen customer engagement or information sites can be optimized to improve the availability of relevant content to the user. In web navigation, the users goal and whether she reached it, is typically unknown. This makes navigation games particularly interesting to researchers, since they capture human navigation towards a known goal and allow building labelled datasets suitable for supervised machine learning models. In this work, we show that a recurrent neural network model can predict game success from a partial click trail without knowledge of the users navigation goal. We evaluate our approach on data from WikiSpeedia and WikiGame, two well known navigation games and achieve an AUC of 86% and 90%, respectively. Furthermore, we show that our model outperforms a baseline that leverages the navigation goal on the WikiSpeedia dataset. A detailed analysis of both datasets with regards to structural and content related properties reveals significant differences in navigation behaviour, which confirms the applicability of our approach to different settings.