Class-Incremental Continual Learning for Multi-View Clustering

Shengsheng Qian, Shengjie Zhang, Dizhan Xue, Huaiwen Zhang, Changsheng Xu

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Class-incremental continual learning, multi-view clustering, data mining, multi-view learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: This paper explores a new problem in data mining (i.e., class-incremental continual learning for multi-view clustering) and proposes an effective method.

Abstract: Multi-view clustering (MVC) aims to explore common semantics for multi-view data and has become an active research topic. However, existing MVC methods focus on learning from static training data and ignore streaming multi-view data with incremental classes, which is frequent in real-world applications given the continually evolving nature of our world. Meanwhile, the existing continual clustering methods only consider single-view data, which cannot effectively mine the semantics of multi-view data. In this paper, we propose a novel Class-incremental Continual Multi-View Clustering (CCMVC) method to handle class-incremental continual learning for multi-view clustering, where multi-view data with incremental semantic classes come sequentially. Our method conducts two iterative optimization phases, i.e., multi-view cluster search and multi-view cluster consolidation, for sequential multi-view training data. In the test, our CCMVC can perform online multi-view clustering for all emerged classes. Firstly, CCMVC learns the common feature space for multi-view data and searches clusters for the incoming data. Secondly, CCMVC harmonizes and consolidates all learned clusters in a unified MVC model with data replay for all emerged classes. In particular, we propose a cross-view synchronous loss to mitigate the asynchronous convergence problem inherent in multi-view continual learning. Extensive experiments on six public MVC datasets reveal the superiority of CCMVC compared with the state-of-the-art methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4582