Inferring user's search activity using interaction logs and gaze data

Abstract

Providing an individual user support regarding the current search activity is crucial for adaptive information retrieval systems but requires a reliable identification of user activities. Such a reliable and robust classification would also help to characterize and analyze complex search behavior of users and enable us to build more advanced information-companion technologies that can support users individually in retrieving and organizing information. In order to develop classification models, in this paper we investigate the validity of interaction features and model structures. We provide a methodology for user models utilizing data from interaction logs and eye tracking to classify the two search activities exploratory search and multitasking search. To identify adequate models and features we use a statistical framework that enables us to select relevant parameters that are useful to understand the current search activity while holding discriminative properties between the activities. The model achieved a classification rate of 89.32%.