Pruned Broad Learning System Based on Sparse Ridge Fusion

<title>Abstract</title> Broad learning system is an emerging method, which has achieved outstanding performance in regression and classification problems. This paper proposes a novel algorithm called Pruned Broad Learning System (PBLS) to reduce model size and improve model interpretability. The proposed PBLS introduces the sparse ridge fusion penalty into BLS, which combines the \({L_1} - norm\) regularizer and second order difference penalty together. The \({L_1} - norm\) regularizer is used to sparse the model by making insignificant output weights toward zero, while second order difference penalty is used to smooth output weights by penalizing the roughness of the model. Then, the unimportant nodes are pruned from the model. The remaining useful nodes are divided in the form of block structure, and their output weights change slowly and smoothly in their respective block structure. The experiments on several commonly used regression data sets are carried out to verify the feasibility of the proposed pruned BLS. The experiment results indicate that pruned BLS can reduce model complexity without loss of prediction accuracy and improve the model interpretability.