Demixing and Topology Identification for Mixed Graph Signals

Graph learning (GL) plays a pivotal role in graph signal processing (GSP) for inferring the underlying signal structure. However, most of the recent works only focus on the single-graph learning situation. This work studies the multiple-graph learning problem for mixed graph signals, and a joint blind signal separation (BSS) and GL algorithm is proposed. To eliminate the ordering uncertainty of BSS, an approach based on maximum graph correlation (MGC) is also put forward such that the unmixed graph signals can be partitioned into a set of clusters, with a learned topology for each cluster. Numerical experiments demonstrate the effectiveness of this algorithm, in both signal separation and graph topology recovery.