Nonnegative matrix factorization algorithms based on the inertial projection neural network

This paper presents two methods for nonnegative matrix factorization based on an inertial projection neural network (IPNN). The first method applies two IPNNs for optimizing one matrix, with the other fixed alternatively, while the second optimizes two matrices simultaneously using a single IPNN. With the proposed methods, different local optimum solutions can be found under the same initial conditions, whereas most traditional methods can only find one local optimum solution. Moreover, experimental results on synthetic data, signal processing, and clustering in real-world data demonstrate the effectiveness and performance of the proposed methods.