Deep Constrained Clustering with Active Learning

Dan HUANG, Ran WEN, Boren DING, Junhua LI

Abstract

Deep semi-supervised clustering approaches, which use supervised data to help the deep neural network acquire cluster-friendly representations, have improved clustering performance and simultaneously increased the semantic value of the clustering results. However, the majority of them cannot utilize both labeled and unlabeled data completely. Furthermore, in these methods, the supervised information is either passively acquired or randomly picked, which may be insufficient, redundant, and even decrease the performance of these models. This paper provides a deep semi-supervised clustering technique with active learning to address the problems mentioned above. The procedure is divided into two sections: model training and data labeling. In the model training section, the paired data is used to train the pseudo-Siamese network, and then the sub networks of the pseudo-Siamese network are fine-tuned using self-training. A new query strategy is devised in the data labeling part, which combines the traditional uncertainty query strategy with the deep Bayesian uncertainty query strategy. Finally, substantial tests are conducted to confirm the utility of the suggested approach on certain real-world data sets. The results of the tests demonstrate that both the suggested method and query strategy are practical.

Keywords

Semi-supervised clustering, Deep learning, Pairwise constraints, Active learning, Pseudo-Siamese network.

View full article