THSLRR: A Low-Rank Subspace Clustering Method Based on Tired Random Walk Similarity and Hypergraph Regularization Constraints

Single-cell RNA sequencing (scRNA-seq) technology furnishes us with a certainly forceful tool for exploring biological mechanisms from the perspective of single-cell. By clustering scRNA-seq data, different types of cells can be effectively distinguished, which is helpful for disease treatment and the discovery of new cell types. Nevertheless, the existing clustering methods still cannot achieve satisfactory results attributed to the complexity of high-dimensional noisy scRNA-seq data. Therefore, we propose a clustering method called Hypergraph regularization sparse low-rank representation with similarity constraint based on tired random walk (THSLRR). Specifically, the sparse low-rank model rebuilds spatial information from a suite of high-dimensional subspaces by mapping data into subspaces, and removes superfluous information and errors in scRNA-seq data. The hypergraph regularization explores the higher-order manifold structure embedded in the scRNA-seq data. Meanwhile, the similarity constraint based on tired random walk can farther upgrade the learning ability and interpretability of the model. Then, the learned similarity matrix could be for spectral clustering, visualization and identification of marker genes. Compared with other advanced methods, the clustering results of the THSLRR method are more robust and accurate.