T-distributed Stochastic Neighbor Network for unsupervised representation learning.

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• Unsupervised representation learning (URL) lacks effective operators for extracting structural information from diverse data types.
• Existing URL methods struggle with preserving local data structures and addressing the "out-of-sample" problem.

Purpose of the Study:

• To propose a novel end-to-end network, T-distributed Stochastic Neighbor Network (TsNet), for unsupervised representation learning.
• To enhance data clustering and visualization through improved representation discrimination and handling of generic data.

Main Methods:

• Developed an adaptive connectivity distribution learning module to construct pairwise graphs preserving local data structure.
• Implemented a T-distributed stochastic neighbor embedding loss function for learning data transformations and improving representation discrimination.
• Incorporated a nonlinear parametric mapping for unsupervised, generalized learning to address the "out-of-sample" issue.

Main Results:

• TsNet significantly outperforms previous unsupervised learning approaches in data visualization and clustering.
• Achieved 74.90% Accuracy (ACC) and 76.56% Normalized Mutual Information (NMI) on single-cell RNA-sequencing (scRNA-seq) datasets.
• Demonstrated an 8% relative improvement over state-of-the-art methods in scRNA-seq clustering.

Conclusions:

• TsNet offers a robust and effective solution for unsupervised representation learning across various data types.
• The proposed method successfully addresses limitations in capturing local structure and handling new data points.
• TsNet shows particular promise for complex biological data analysis, such as scRNA-seq data clustering.