Multi-View Diffusion Proces for Spectral Clustering and Image Retrieval
Date:
The flowchart of MVD
In this talk, I introduce Multi-View Diffusion (MVD) [slides, code, paper], a novel approach to multi-view graph learning that combines weight learning with graph learning in an alternating optimization framework. This method excels in situations without prior data distribution knowledge, creating a unified affinity graph from diverse data sources.
Our novel fusion-and-diffusion strategy merges multiple affinity graphs through a weight learning scheme, leading to a consensus foundation for further diffusion. The core of this approach is a multi-view diffusion process that enhances pairwise affinities by spreading them across tensor product graphs, capturing higher-order contextual information.
This method is robust, avoiding the typical constraints of relying on initial graph quality or assuming a common subspace. Tested on 16 real-world datasets, it outperformed leading techniques in 13 cases, proving its effectiveness in image retrieval and clustering tasks. This presentation will highlight the key innovations and practical impacts of this approach.