Learning bipartite graphs from spectral templates

Graph learning is crucial for understanding the relationship between data components. Signal processing-based graph learning algorithms are designed for specific signal models. This work investigates the problem of learning bipartite graphs given arbitrarily ordered spectral templates or graph eigenvectors. Starting from the spectral templates, the proposed algorithm identifies the vertex groups of the bipartite graph. Experiments conducted on three different types of synthetic datasets demonstrate that the proposed bipartite graph learning algorithms outperform structure-blind learning techniques across various signal-to-noise (SNR) regimes. Our algorithm leverages the spectral signatures of a bipartite graph, specifically the structure of the graph’s eigenvectors.