Deep Clustering with Uniform Quasi-low-rank Hypersphere Embedding
Keywords: unsupervised learning, representation learning, deep clustering
TL;DR: A novel unsupervised representation learning framework is proposed to learn uniform quasi-low-rank hypersphere embedding for clustering.
Abstract: With the powerful representation ability of neural networks, deep clustering (DC) has been widely studied in machine learning communities. However, current research on DC has rarely laid emphasis on the inter-cluster representation structures, i.e. ignoring the performance degradation caused by the low uncorrelation between different clusters. To tackle this problem, a Uniform quasi-Low-rank Hypersphere Embedding based DC (ULHE-DC) method is proposed herein, which promotes learning an inter-cluster uniform and intra-cluster compact representation in a novel geometric manner. Specifically, clusters are uniformly distributed on a unit hypersphere via minimizing the hyperspherical energy of the centroids, and the embeddings belonging to the same cluster are simultaneously collapsed to a quasi-low-rank subspace through intra-cluster correlation maximization. Additionally, a pre-training based optimization scheme is proposed, in which an auto-encoder (AE) is pre-trained and the parameters of the encoder of AE are inherited to initialize the feature extractor for clustering, aiming at engaging the model learning cluster-oriented representation more efficiently. Experimental results validate the strong competitiveness of the proposed method, compared with several state-of-the-art (SOTA) benchmarks.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 10499
Loading