Reproduce “Domain Adaptation Principal Component Analysis: base linear method for learning with out-of-distribution data” on OpenResearch

About this paper

Evgeny M Mirkes, Jonathan Bac, Aziz Fouché, Sergey V. Stasenko, Andrei Zinovyev, Alexander N. Gorban

Researchers from University of Leicester, Institut Curie, and Lobachevsky University introduce Domain Adaptation Principal Component Analysis (DAPCA), a linear method for supervised domain adaptation that aligns data distributions across domains while performing dimensionality reduction. DAPCA demonstrated robust performance in aligning data distributions for classification tasks on diverse datasets, including single-cell omics, and proved computationally efficient compared to deep learning alternatives.

Key insights

• Explicitly defining attraction between similar points across domains, based on kNN and source labels, effectively mitigates domain shift while preserving discriminative features.
• A linear, optimization-based approach can serve as a robust, interpretable, and computationally efficient baseline or preprocessing step, offering advantages over complex deep learning methods.
• The iterative nature of DAPCA, where kNN relationships are updated in the projected space, ensures convergence and provides a robust solution for aligning data distributions.

Summary by alphaXiv.