Personalised Session Difficulty Prediction in an Online Academic Search Engine

摘要

Search sessions consist of multiple user-system interactions. As a user-oriented measure for the difficulty of a session, we regard the time needed for finding the next relevant document (TTR). In this study, we analyse the search log of an academic search engine, focusing on the user interaction data without regarding the actual content. After observing a user for a short time, we predict the TTR for the remainder of the session. In addition to standard machine learning methods for numeric prediction, we investigate a new approach based on an ensemble of Markov models. Both types of methods yield similar performance. However, when we personalise the Markov models by adapting their parameters to the current user, this leads to significant improvements.