Reproduce “A Trainable Centrality Framework for Modern Data” on OpenResearch

About this paper

Minh Duc Vu, Mingshuo Liu, Doudou Zhou

Researchers at the National University of Singapore and the University of California, Davis, introduce Fused Unified centrality Score Estimation (FUSE), a neural framework designed to measure data centrality by combining global and local signals. This framework operates on arbitrary data representations, offering significantly faster inference and improved adaptability over classical depth notions for modern, high-dimensional data.

Key insights

• Centrality can be effectively learned in a self-supervised manner by combining preference learning (Bradley-Terry model) for global typicality and generative modeling (Denoising Score Matching) for local density estimation.
• The framework's homotopy parameter 't' offers an interpretable and continuous mechanism to explore multi-scale data geometry, smoothly transitioning between broad global views and fine-grained local structures.
• By operating on fixed, pre-trained feature representations, FUSE extends the utility of data centrality to diverse high-dimensional and non-Euclidean data modalities, overcoming a key limitation of classical methods.

Summary by alphaXiv.