An optimization framework for hierarchical clustering

Hierarchical clustering is a fundamental problem in computational biology, with popular greedy heuristics such as average linkage dating back to the 1950s but no well-defined objective. Recently, a combinatorial optimization criterion for the problem was suggested by Dasgupta. While minimizing this criterion is NP-hard, the popular average linkage method serves as a strong baseline. Nevertheless, its myopic, greedy nature frequently leads to structurally suboptimal hierarchies.

To remedy this, we introduce a novel average-linkage-based clustering approach that combines local and global considerations by generating multiple views of the input data and learning how to blend them into an integrated similarity measure. We demonstrate that our method, DOMUS, consistently outperforms strong baselines, including a beam search heuristic, on a wide range of synthetic and classic benchmark datasets. Furthermore, we validate its real-world applicability through a rigorous benchmark on single-cell RNA sequencing data, where it compares favorably with the state-of-the-art HiDeF algorithm.

The DOMUS framework is implemented in Python and freely available at https://github.com/GalGilad/DOMUS.