Affiliations: [a] Faculty of Computer Science, Dalhousie University, Halifax, Nova Scotia, Canada | [b] School of Management, Dalhousie University, Halifax, Nova Scotia, Canada
Note: [1] Short version of this paper appears in [13].
Abstract: Unsupervised document clustering groups documents into clusters without any user effort. However, the clusters produced are often found not in accord with user's perception of the document collection. In this paper we describe a novel framework and explore whether clustering performance can be improved by including user supervision at the feature level. Unlike existing semi-supervised clustering methods, which ask the user to label documents, this framework interactively asks the user to label features. The proposed method ranks all features based on the recent clusters using cluster-based feature selection and presents a list of highly ranked features to the user for labeling. The feature set for the next clustering iteration includes both features accepted by the user and other highly ranked features. The experimental results on several real datasets demonstrate that the feature set obtained using the new interactive framework can produce clusters that better match the user's expectations compared with the unsupervised version of the methods. Moreover, we quantify and evaluate the effect of reweighting previously accepted features and of user effort. Different underlying clustering algorithms such as K Means and Multinomial Naïve Bayes model are demonstrated to perform very well with the newly proposed framework.
