Multi domain clustering improves grouping of diverse data types
Multi-Domain Clustering via Measure Quantization
Machine Learning
Summary
Clustering means grouping similar data points together, and doing this across different types of data is a challenge. The authors propose a method that finds common representative points across multiple data domains by minimizing differences measured by advanced distance metrics. This method assigns data points to these shared groups either by closeness or by a method that considers all points together, making grouping more accurate. They also use a strategy that breaks data into manageable batches, helping to handle large datasets efficiently.
What this means in practice
- •For data engineers: Cluster large, mixed-type datasets from different sensors or modalities with a single shared representation for improved analysis consistency.
- •For machine learning platform developers: Integrate scalable multi-domain clustering into platforms to enable unified grouping and feature extraction across image, audio, and sensor inputs.
Authors
Rafael Pereira Eufrazio, Eduardo Fernandes Montesuma, Charles Casimiro Cavalcante
Abstract
Clustering is a fundamental task in data analysis, typically addressed through centroid-based methods such as K-means. In this work, we present a general framework for multi-domain clustering via measure quantization: given samples from multiple domains, we learn a shared set of cluster prototypes by minimizing a probability metric, such as the Sinkhorn divergence or the Maximum Mean Discrepancy, between each domain's probability measure and the measure of prototypes. Data points are then assigned to clusters either via nearest centroid, or via optimal transport, a collaborative strategy that couples all samples within a domain. A mini-batch optimization strategy makes both fitting and assignment scalable, reducing memory and computational cost while preserving clustering performance. Experimental results on 5 multi-domain benchmarks spanning image, audio and sensor data show that our Sinkhorn-based method consistently outperforms classical and multi-domain clustering baselines, and that this advantage persists when scaling to hundreds of thousands of samples.