From User Query to User Model and Back: Adaptive Relevance-Based Visualization for Information Foraging Jae-wook Ahn and Peter Brusilovsky* School of Information Sciences, University of Pittsburgh ABSTRACT In the context of DARPA GALE project, out team attempts to augment the set of information foraging tools available to an intelligence analyst with adaptive information filtering. The adaptive filtering engine of ROSETTA system developed by our joint project team augments traditional query-based ranking provided by a search component with a user profile-based ordering of information passages. However, professional analysts were interested to have more control over the ranking by mediating between two extremes - query-based and profile-based ranking. To answer this need we developed an adaptive relevance-based visual exploration component for ROSETTA. This component re-purposes the VIBE visualization approach developed earlier in our School and extends it with a range of tools to support analysts in the process of query to profile mediation. This paper presents the rationale and the functionality of our visual exploration component and reports the results of its preliminary evaluation. CR Categories and Subject Descriptors: H.3.1 [Content Analysis and Indexing]: Indexing method; H.3.3 [Information Search and Retrieval]: Information filtering; Relevance feedback; H.3.5 [Online Information Services]: Web-based services; H.5.2 [User Interfaces]: Graphical user interfaces (GUI). Additional Keywords: Intelligence Analysis, Visualization, VIBE, Ranked List Fusion, Query, User profile 1 INTRODUCTION Powerful information access tools are important for intelligence analysts at information foraging stage of their work [9]. During the information foraging stage the analysts use their domain and search knowledge to collect potential useful information from documents in various media and sources. In current generation of tools for information analysts the set of information access tools is typically limited to traditional information retrieval tools. In response to analysts’ queries these tools provide a list of links to information resources ranked by their relevance to a query. In the context of DARPA GALE (http://www.darpa.mil/ipto/ programs/gale) project, out team attempts to augment the set of information foraging tools available to an intelligence analyst with adaptive information filtering [3]. An adaptive filtering engine collects potential useful information about users’ interests, preference, and knowledge from either explicit feedbacks from users’ relevance judgments or implicit feedbacks by observing the users’ search and browsing activities, it then uses such information to predict and recommend potential relevant information to user. In several relevant contexts, such as news reading [2] or TV program selection [7] personalized information filtering tools already demonstrated their value, however there are still very few attempts to apply these tools in the context of intelligence analysis [11][12]. Figure 1 ROSETTA System Interface The ROSETTA system (Figure 1) developed by our joint project team includes both search and personalized filtering tool. The core of this tool is personalized filtering engine CAFÉ (Carnegie Mellon Adaptive Filtering Engine), which collects various kinds of user implicit feedback and builds a profile of user interests. CAFÉ processes information retrieved by the search component of ROSETTA and re-ranks it according to the user profile. In two rounds of extensive evaluation of ROSETTA with professional intelligence analysts, CAFÉ was highly praised. However, the users also indicated that they want to have more control over the performance of the engine. In particular, the users were interested to have more control over the ranking by mediating between two currently available extremes – query-based ranking in the search component and profile-based ranking in the filtering component. The problem of “fusing” query-based ranking and profile-based ranking is not new. The traditional solution of this problem, which is applied in several adaptive search systems [6], is to select a fixed mediation point α between 0 and 1 and to produce a personalized rank by fusing query-based and profile-based ranks with coefficients α and (1-α). By manipulating α, the system designers can give more priority to documents similar to the query or documents similar to the profile. However, this solution is not providing the analysts with the requested flexibility. This paper presents a more flexible approach to “fusing” query-based and profile-based ranking. The idea of this approach is to allow the * address: 135 N. Bellefield Ave, Pittsburgh PA 15256, USA e-mail: {jaa38, peterb}@pitt.eduanalysts to decide dynamically whether they are interested in documents, which are closer to the query or documents which are closer to the user profile navigating all the way from the query to the user profile and back. The core component of our approach is relevance-based visualization originally implemented in VIBE [8]. VIBE is known as an excellent tool for visual query results analysis. It allows the user to explore connection between search results and query terms, picking, for example, a subset of results that is more relevant to a specific query term or a group of terms. Our visual analysis tool applies relevance-based visualization to help the user to mediate between the query terms and terms from the user profile. The application of user profile makes relevance-based visualization adaptive. The results of visualization are different for different users issuing the same query and even different at different time for the same user as long as the interests of this user represented in the user profile evolve. This paper presents our implementation of VIBE for adaptive relevance-based visualization, stresses several features that are critical for this type of visualization, and presents some evaluation results. 2 FUSING QUERY AND PROFILE-BASED RANKINGS WITH VIBE VIBE (Visual Information Browsing Environment) was originally developed in Molde College in Norway and the School of Information Sciences at the University of Pittsburgh [8]. It is a document visualization tool which supports POI (Point of Interest) based browsing. POIs represent key concepts or keywords and are presented as user draggable icons on the screen. The documents are located