Meta-neuron learning based spiking neural classifier with time-varying weight model for credit scoring problem

摘要

This paper presents a meta-neuron learning-based spiking neural classifier with a time-varying weight model (MeST). MeST is developed to handle the class imbalance in classification problems without any data preprocessing methods. Meta-neuron based learning algorithm in MeST uses normalized postsynaptic potentials (global information) and weight of the connection (local information) to determine the sensitivity modulation factor. This modulation factor determines the proportion of the weight update for a given set of presynaptic spikes. The weight update is then embedded in a Gaussian function to determine the time-varying weight update. The centre of the time-varying Gaussian function is determined by the presynaptic spike times. MeST is demonstrated on 10 benchmark datasets from the University of California, Irvine California machine learning repository and then applied to solve credit scoring using three real-world datasets. Performance studies show that the generalization ability of MeST is better than other spiking neural networks with constant weight model, despite having a simple architecture. Furthermore, compared to other non-spiking shallow machine learning classifiers, MeST is a slightly better model for classification using highly imbalanced datasets. This indicates the learnability of a stand-alone classifier on an imbalanced dataset can be increased by using time-varying weights.