A new semi-supervised fuzzy clustering method based on latent representation learning and information fusion

Fuzzy clustering is a simple but efficient clustering method, which aims to deal with ambiguous and overlapping data classification boundaries and provide detailed membership degree information. However, the complex structure of high-dimensional data in the real world easily causes the inefficiency of widely used distance metrics. In addition, traditional fuzzy clustering methods cannot utilize supervision information to guide membership degree matrix learning, which limits clustering performance. In this paper, we propose a new semi-supervised fuzzy clustering method based on latent representation learning and information fusion. Specifically, our method utilizes a deep autoencoder to learn the underlying low-dimensional structure of data, thereby mitigating the impact of redundant features. Moreover, a semi-supervised constraint term based on information fusion is designed to make full use of prior knowledge to supervise the clustering process. On the one hand, a new constraint distance is proposed by leveraging pairwise constraint information to re-evaluate the membership degree of data. On the other hand, the semi-supervised constraint term is constructed based on label information to guide the membership degree matrix learning. Comprehensive experiments on a variety of standard datasets show that our method achieves better performance compared with state-of-the-art baseline methods, demonstrating the effectiveness of the proposed method in fuzzy clustering.