Structure-Preserving Multi-View Embedding Using Gromov-Wasserstein Optimal Transport

A team of researchers has introduced a novel framework for multi-view data analysis, a common challenge in machine learning where the same object is described by different data types (e.g., an image, a text description, and a 3D scan). Their paper, "Structure-Preserving Multi-View Embedding Using Gromov-Wasserstein Optimal Transport," presents two core methods: Mean-GWMDS and Multi-GWMDS. Both leverage Gromov-Wasserstein (GW) optimal transport, a mathematical framework for comparing data structures by measuring the distance between their internal relationships, rather than forcing a direct alignment of features. This is a significant shift from classical approaches that rely on simple concatenation or assume data views can be linearly aligned, which often fails with complex, heterogeneous data.

The proposed Mean-GWMDS strategy works by averaging the distance matrices from each data view and then applying GW-based multidimensional scaling to find a single, representative low-dimensional embedding. The alternative, Multi-GWMDS, generates multiple candidate embeddings through GW alignment and then selects the most geometry-consistent one. Experiments demonstrated that these methods effectively preserve the intrinsic relational structure across different views on both synthetic datasets and real-world applications. The work, currently under review for the journal Signal Processing, positions GW optimal transport as a flexible and principled foundation for building more robust AI systems that can fuse information from diverse and misaligned sources.