Self-supervised Adaptive Kernel Nonnegative Matrix Factorization.

Furong Deng,Yang Zhao,Jihong Pei,Xuan Yang

DOI: https://doi.org/10.1007/978-3-031-18907-4_13

2022-01-01

Abstract:The kernel function to imply nonlinear mapping is crucial for kernel nonnegative matrix factorizations (KNMF), which determines whether the model can properly represent the data. However, the analysis and adaptive learning of parameters in the kernel function are lacking in existing methods. The kernel parameters of most methods are randomly selected, and the purpose and motivation of kernel function construction is weak. To address this problem, a self-supervised adaptive kernel nonnegative matrix factorization (SAKNMF) is proposed to handle image recognition tasks. Kernel methods are used to achieve non-linearization. Firstly, the data is devided into multiple clusters by fuzzy clustering as self-supervised information. Then, according to the self-supervision information, the discriminative constraints that can adaptively learn the kernel parameters are constructed in the objective function. Based on the new objective function, the Gaussian kernel function parameters, basis matrix and coefficient matrix are jointly optimized. The kernel function constructed by the learned parameters enables the nonlinear mapped basis images to cover the mapped samples in the same cluster and distinguish other different clusters. So that the model can finally learn features that reflect the sample difference and have good generalization. Experiments show that our method can learn appropriate kernel width in different datasets, and the recognition accuracy is better than the state-of-the-art methods.