Interpretable Constructive Algorithm for Random Weight Neural Networks

In this paper, an interpretable construction method (IC) with geometric information is proposed to address a significant drawback of incremental random weight neural networks (IRWNNs), which is the difficulty in interpreting the black-box process of hidden parameter selection.The IC utilises geometric relationships to randomly assign hidden parameters, which improves interpretability. In addition, IC employs a node pooling strategy to select the nodes that will both facilitate network convergence. The article also demonstrates the general approximation properties of IC and presents a lightweight version tailored for large-scale data modelling tasks. Experimental results on six benchmark datasets and one numerical simulation dataset demonstrate the superior performance of IC compared to other constructive algorithms in terms of modelling speed, accuracy and network structure. In addition, the effectiveness of IC is validated by two real-world industrial applications.