Dealing with Concept Drift in Exploratory Search: An Interactive Bayesian Approach

In exploratory search, when the user formulates a query iteratively through relevance feedback, it is likely that the feedback given earlier requires adjustment later on. The main reason for this is that the user learns while searching, which causes changes in the relevance of items and features as estimated by the user -- a phenomenon known as {it concept drift}. It might be helpful for the user to see the recent history of her feedback and get suggestions from the system about the accuracy of that feedback. In this paper we present a timeline interface that visualizes the feedback history, and a Bayesian regression model that can estimate jointly the user's current interests and the accuracy of each user feedback. We demonstrate that the user model can improve retrieval performance over a baseline model that does not estimate accuracy of user feedback. Furthermore, we show that the new interface provides usability improvements, which leads to the users interacting more with it.