计算机科学
聚类分析
人工智能
特征(语言学)
数据挖掘
模式识别(心理学)
模糊聚类
领域(数学)
口译(哲学)
特征选择
作者
Zhenqiu Shu,Rong Dong,Tianyang Xu,Hongbin Wang,Zhengtao Yu
标识
DOI:10.1109/tcbbio.2026.3731145
摘要
Single-cell RNA sequencing (scRNA-seq) enables high-resolution analysis of cellular heterogeneity. Accurate scRNA-seq data clustering is crucial for cell type identification and subpopulation discovery. However, existing clustering methods are frequently affected by erroneous pseudo-labels generated through the static threshold, thus progressively accumulating errors during training. To address these challenges, in this paper, we propose a novel curriculum-guided contrastive co-training (scCGC$^{2}$T) framework for scRNA-seq data clustering. It designs a dual-branch architecture for cross-view contrastive co-training, thereby strengthening its representation ability for scRNA-seq data. Additionally, a progressive curriculum learning strategy is introduced to filter pseudo-labels dynamically. Specifically, high-confidence samples using the static threshold are applied to initial training, followed by guidance based on dynamic thresholding. Low-confidence samples are gradually incorporated into the training process, thereby mitigating the noise interference problem. Therefore, the proposed scCGC$^{2}$T method effectively mitigates error accumulation in the early training stage and greatly enhances the discriminative capability for complex cell subpopulations. Extensive experiments on several benchmarks demonstrate that the proposed method achieves superior clustering performance and strong generalization on scRNA-seq data clustering tasks.
科研通智能强力驱动
Strongly Powered by AbleSci AI