An Incremental Broad Learning Approach for Semi-Supervised Classification

Abstract

Broad Learning System (BLS) is a fast and accurate supervised learning method without deep structure. However, the classifier trained by BLS cannot achieve expected accuracy if the labeled data are insufficient. In this paper, we develop an Incremental Semi-supervised Broad Learning method (ISBL) based on BLS to classify a partially labeled dataset, which applies manifold regularization to explore the underlying data distribution and improve accuracy. ISBL applies to scenarios where data is generated over time. While new patterns are added to the learning algorithm, the proposed method updates the classification model sustainability without retraining. By comparing with original BLS and other semi-supervised classification techniques on various datasets with different dimensions, we verified that ISBL outperforms these methods on accuracy. Experimental results demonstrate that ISBL utilizes unlabeled data effectively and achieves high accuracy. Meanwhile, the incremental learning method reduces the learning time and storage of historical data.