User Centred Evaluation of A Recommendation Based Image Browsing System

In this paper, we introduce a novel approach to recommend images by mining user interactions based on implicit feedback of user browsing. The underlying hypothesis is that the interaction implicitly indicates the interests of the users for meeting practical image retrieval tasks. The algorithm mines interaction data and also low-level content of the clicked images to choose diverse images by clustering heterogeneous features. A user-centred, task-oriented, comparative evaluation was undertaken to verify the validity of our approach where two versions of systems { one set up to enable diverse image recommendation { the other allowing browsing only { were compared. Use was made of the two systems by users in simulated work task situations and quantitative and qualitative data collected as indicators of recommendation results and the levels of user's satisfaction. The responses from the users indicate that they nd the more diverse recommendation highly useful.