Visualizing User Model in Exploratory Search Tasks

We present our ongoing work on an interactive exploratory information retrieval system designed to explain to the user the reasons for its predictions and help the user to build a mental model to predict how the system will behave. Users indicate which documents interest them and reinforcement learning is used to model the user by allowing the system to trade o↵ between exploration and exploitation. In most interactive systems, the user has a concept in mind that the system is trying to learn while the user interacts with the system, e.g. a particular genre of music or specific types of documents. Many of such systems rely on machine learning techniques to help the system to identify the concept that the user has in mind. In recent years, there has been a growing interest in applying various reinforcement learning (RL) techniques in exploratory information retrieval systems. RL allows the system to trade between exploitation (moving towards more specific subtopics) and exploration (going towards alternative topics), and has been shown to be helpful in information retrieval, in particular in exploratory search systems, where the final search goal maybe vague [2]. However, experimental results show that users of such information retrieval often feel that they are not fully in control of the system and sometimes they do not fully understand why a given set of results was returned to them by the information retrieval system [2]. In this paper, we address the issue of how to assist the user in understanding how the exploration – exploitation trade-o↵ in a RL based information retrieval system a↵ects their search results. We describe a prototype document retrieval system based on the exploration – exploitation trade-o↵ to support exploratory search. Appearing in Proceedings of the 4 Workshop on Machine Learning for Interactive Systems (MLIS) 2015, Lille, France. JMLR: W&CP volume 40. Copyright 2015 by the authors. Figure 1: System interface with the query ”big data”. Our ongoing evaluation of the system indicates that the additional visualization allows the user to better understand the workings of the algorithm used in the system and obtain better search results.