Interaction Design for Constraints Selection in Interactive Constrained Clustering

• Published in: Transactions of the Japanese Society for Artificial Intelligence Vol. 29; no. 2; pp. 259 - 267
• Main Authors: Yamada, Seiji, Okabe, Masayuki, Mizukami, Junki
• Format: Journal Article
• Language: English; Japanese
• Published: Tokyo The Japanese Society for Artificial Intelligence 2014; Japan Science and Technology Agency
• Subjects: human active learning; interaction design; interactive constrained clustering
• ISSN: 1346-0714; 1346-8030
• DOI: 10.1527/tjsai.29.259

In this paper, we describe a novel GUI for human constraints selection in interactive constrained clustering. Clustering is one of the most popular data mining technologies, and in particular, interactive constrained clustering that uses constraints obtained from a human is promising for practical applications. However, some constraints are not effective in constrained clustering, and the cognitive load necessary for a user to give constraints is high in interactive clustering, thus we need to provide a mechanism in which a user can easily select only effective constraints for an interactive constrained clustering system. Since the constraints are considered to be training data for classification learning, traditional computational active learning like uncertainty sampling might be useful in non-interactive constrained clustering. For interactive constrained clustering, we should use human constraints selection. Thus we build a GUI for human constraints selection in interactive constrained clustering. Our GUI has two functions to effectively derive human constraints selection; exposing the effects of given constraints and providing multiple viewpoints. As the first function, we propose a GUI that can expose the effects of given constraints by emphasizing them at the clustered results. The second function can provide multiple viewpoints that a user can flexibly change to derive human constraints selection. We fully implemented an interactive constrained clustering system with the proposed GUI. We also conducted an evaluation experiment on image clustering with participants, and obtained results to support the effectiveness of our approach to derive human constraints selection.