Evolutionary Image Co-clustering with User Feedbacks

Traditional image clustering systems are primarily based on visual and/or textual features. Such algorithms commonly suffer from the problem of semantic gap. General approach to overcome this problem is to incorporate user feedback. However, data extracted from certain domains like social networks, web-blogs etc. is evolving in nature. In other words, data collected from such domains over small period of time interval exhibits high similarity over data instances and features, which can be effectively used to optimize data clustering since change in clustering is gradual. In this paper, we propose EHCC (Evolutionary starstructured Heterogeneous Co-Clustering) algorithm for image co-clustering. Our algorithm incorporates user provided feedbacks over period of time to guide coclustering process. We incorporate user provided feedback in terms of image similarity logs over period of time to augment relational matrix obtained from low level features (color and textual features) extracted from images. Through an iterative algorithm, we trifactorize new relational matrix to obtain image clusters. Through extensive experiments on image data sets, we demonstrate effectiveness and efficiency of our proposed algorithm. keywords: image, clustering, user, feedback, matrix, factorization.