Explainable data generation from observed small samples by matrix tri-factorization

Data generation on small samples has exhibited great potential on real-world biomedical scenarios, especially on extreme cases such as early-stage or emerging biomedical conditions (e.g. Patient Zero of a brand-new disease with only one sample) because of insufficient sample information. Herein, we design an explainable data generation approach under extreme sample scarcity by exploiting inherent low-dimensional principles contained in observed high-dimensional data with comprehensive information, named by Explainable Data Generation (EDG) based on matrix tri-factorization. EDG aims to represent observed small high-dimensional samples by low-dimensional latent representation with a larger rather than equal sample size. This leads to explainable sample generation, due to transparency and interpretable process of decomposition and reconstruction. In experiments, EDG is tested on four tasks of classification, clustering, feature selection and tipping point prediction on simulated and real datasets (single-cell RNA-seq / ATAC-seq data) under extreme sample scarcity, with results demonstrating the superiority of EDG. Here the authors propose an explainable data generation approach called EDG that operates under extreme sample scarcity by exploiting inherent low-dimensional principles contained in the observed high-dimensional data via matrix tri-factorization.