Dynamic Hypergraph Regularized Broad Learning System for Image Classification

As a novel alternative to deep neural networks, the broad learning system (BLS) has exhibited outstanding performance in many machine learning tasks. The unique flat network structure enables BLS to extract the non-linear characteristics of the data. And the weights of the hidden layer are randomly generated, allowing BLS to provide rapid training for large-scale data. However, BLS focuses on approximating the target data and ignores the inherent geometric structure of the data. Graph structure can reflect the latent structure between samples of a dataset. Hypergraph learning has superior performance in modeling high-order relationships compared to simple graph-based learning methods that can only model pairwise relationships of data. In this paper, we propose a novel extension of the standard BLS. The proposed dynamic hypergraph regularized broad learning system (DHGBLS) incorporates hypergraph learning in the optimization process. And dynamic optimization is established to learn the hypergraph and the network’s output weights simultaneously. In this way, the effects of various hyperedges can be automatically modulated. Experimental results on three popular datasets show the superiority of the proposed method over the standard BLS and other state-of-art classification methods.