Semi-supervised Clustering Algorithm Based on L1 Regularization and Extended Pairwise Constraints

Clustering algorithm plays an important role in recommendation system, data analysis, market segmentation and other fields. Their primary objective is to grow up similar samples into the same clusters while separating dissimilar samples into distinct clusters. With the rapid development of deep learning, deep clustering integrates the strong representational ability of deep learning into clustering tasks and achieves outstanding performance. Currently, most clustering algorithms are unsupervised. However, many times datasets contain not only unlabeled data but also limited relational information. It is possible to significantly enhance the clustering performance by using this prior information effectively. Thus, this research focuses on semi-supervised clustering for data involving the extended pairwise constraints. Specifically, this paper adopts the following techniques to improve deep clustering effectiveness: (1) Use partial samples with pairwise information to extend the whole dataset, which makes full use of the relationship information between samples. (2) Add the L1 norm to the loss function, allowing for feature sparsity to enhance model generalization. (3) Update cluster centers with KL divergence to introduce more information during center adjustments. Ultimately, experimental results from five datasets demonstrate significant enhancements in clustering performance achieved by the semi-supervised clustering algorithm proposed in this study.