Adaptive Symmetric NMF for graph clustering

Organizing data into clusters is a key task for data compression and classication. In this paper we consider the case where the data are points belonging to a linear space, whose distance is measured through the Euclidean norm. A symmetric modeling of the graph clustering problem is addressed and an algorithm is proposed, based on NMF (nonnegative matrix factorization) techniques applied to a penalized nonsymmetric minimization problem. The solution depends on several parameters, whose choice is crucial. To overcome this difficulty, we suggest a heuristic approach which detects the best parameter values in an adaptive way. Extensive experimentation shows that the proposed algorithm is effective.