Broad learning system with manifold regularized sparse features for semi-supervised classification

Abstract

Broad learning system (BLS) is an efficient neural network, and is proven to be effective in fields like remote sensing, fault diagnosis, etc. As a critical branch of BLS, semi-supervised BLS has drawn increasing attention. Exploiting the information within additional unlabeled instances is key to semi-supervised learning. Studies have shown that incorporating this information into the feature nodes is a good way to implement semi-supervised BLS. However, the existing methods could not retain the sparsity of feature nodes. Besides that, these methods become computation consuming when dealing with the large scale datasets. To address these problems, a broad learning system with manifold regularized sparse features (BLS-MS) is proposed. We first propose a manifold regularized sparse autoencoder based on extreme learning machine (MS-ELM-AE) for feature mapping. Then, a subset training approach is introduced to alleviate the efficiency decline caused by large data size. Finally, the proposed BLS-MS is further modified to utilize the discriminant information of labeled data, namely discriminative BLS-MS (DBLS-MS). The proposed methods have been evaluated on 14 datasets. Experiment results have demonstrated both the effectiveness and the efficiency of proposed methods.