Document Clustering With Dual Supervision Through Feature Reweighting

Abstract

Traditional semi‐supervised clustering uses only limited user supervision in the form of instance seeds for clusters and pairwise instance constraints to aid unsupervised clustering. However, user supervision can also be provided in alternative forms for document clustering, such as labeling a feature by indicating whether it discriminates among clusters. This article thus fills this void by enhancing traditional semi‐supervised clustering with feature supervision, which asks the user to label discriminating features during defining (labeling) the instance seeds or pairwise instance constraints. Various types of semi‐supervised clustering algorithms were explored with feature supervision. Our experimental results on several real‐world data sets demonstrate that augmenting the instance‐level supervision with feature‐level supervision can significantly improve document clustering performance.

Interactive Clustering

“…Arguably, human expert judgment is needed to balance these requirements by steering the clustering process and evaluating the results [54]. In the same context, in Bekkerman et al [16], Choo et al [30], Corrêa et al [34], El-Assady et al [42], Hoque and Carenini [53], Hu et al [56][57][58], the interaction supports generating clusters that are based on user's domain knowledge in accordance with their understanding to get clusters that fit user's expectations well. In Awasthi and Zadeh [10], the following scenario is given as motivation: "Consider documents representing news articles that could be clustered as {politics, sports, entertainment, other}; however, perhaps the user would like these articles to be clustered into {news articles, opinion pieces}" ([10], p. 1).…”

“…In several cases, user feedback involves updating the weights, either different data instances or different features [34,[56][57][58]. For example, Babaee et al [11] discovers the semantic structure of synthetic-aperture radar (SAR) image collections by letting users give higher weights to the most influential image in each cluster.…”

“…More recently, in the INCREMENT approach (Mitchell [73]) selected instances from the sub-clusters are presented to a user for labeling to improve query efficiency. In studies concerned with clustering documents, individual words are considered to be features, and sometimes the user adapts the the word's weight based on the iterative results [56][57][58]97]. In Bekkerman et al [16], the users are kept active throughout the clustering process by selecting suitable features for the task-specific document representation.…”

“…Regarding to clustering text documents, the methods in Hu et al [56][57][58] present a subset of top-ranked features to the user. A user can tag features as "accepted" if they are believed to be useful for discriminating clusters for every iteration.…”

“…Both accuracy and normalized mutual information (NMI) appear very often as evaluation metrics. For instance, Babaee et al [11] uses accuracy (percentage of correct labels based on the prior annotations) and NMI (0-1, which decides whether two clusters are dependent or independent); Hu et al [56][57][58] uses accuracy and NMI as well. In Andrienko et al [7], the ratio of false negatives and false positives is used in addition to the subjective assessment for evaluating cluster cohesion.…”