New Theory Reveals When Transfer Learning Boosts High-Dimensional Clustering

Chakraborty and Nandy tackle the fundamental problem of clustering in high dimensions when auxiliary source datasets are available. Working under a two-community Gaussian mixture model, they introduce a transfer-assisted clustering procedure that is minimax-optimal—meaning it achieves the best possible worst-case error rate. The key contribution is a precise characterization, up to logarithmic factors, of the phase transition where source data begins to improve target clustering. This depends on the signal-to-noise ratios, sample sizes, ambient dimension, and the geometric alignment between the target and source cluster means. The theory also provides an adaptive mechanism to decide whether to use the source data or rely solely on the target.

The authors extend their framework to handle multiple communities and multiple source datasets, making it broadly applicable. Extensive simulations confirm the theoretical predictions. As a real-world demonstration, they analyze a human lung single-cell RNA-sequencing atlas, showing that their method effectively borrows information from related cell-type clusters to improve the identification of rare or noisy cell populations. This work bridges a critical gap between transfer learning theory and practical high-dimensional clustering, offering data scientists a principled way to leverage related datasets for more accurate unsupervised learning.