Broad learning system based on adaptive lasso

Abstract

Broad learning system is an emerging method, which has achieved outstanding performance in regression and classification problems. This paper proposes a novel algorithm called Broad Learning System based on adaptive lasso (AdL1-BLS) to reduce model size and improve model generalization performance. The proposed AdL1-BLS introduces the weighted L1-norm penalty into BLS, adaptive lasso is used to assign different adaptive coefficients to each output weight of BLS. Because of the effect of adaptive coefficients, the output weight of BLS gets varying degrees of compression rates, which avoids the problem of over-sparse network structure and reduces the prediction error of the network, which better solve the BLS over-fitting problem, and improves the generalization of BLS under the premise of ensuring the sparseness of the network. Finally, the experiments on several commonly used regression data sets are carried out to verify the feasibility of the proposed AdL1-BLS. The experiment results indicate that AdL1-BLS can reduce model complexity without loss of prediction accuracy and improve the model interpretability.