A model-based deep learning method (SLoG-Net) that localizes the sparse source nodes of a diffusion process on a known graph when the diffusion filter is unknown. It casts the problem as blind deconvolution of graph signals, uses filter invertibility to obtain a convex ADMM solver, then unrolls and truncates the ADMM iterations into a trainable neural network whose step-size and penalty parameters are learned end-to-end, yielding interpretable, parameter-efficient inference that matches the iterative solver at a fraction of the runtime.
SLoG-Net is a model-based deep learning architecture for source localization on graphs: recovering the few source nodes that injected a signal which then diffused over a known network, when the diffu…