Artificial Immune System: An Effective Way to Reduce Model Overfitting

Abstract

Artificial immune system (AIS) algorithms have been successfully applied in the domain of supervised learning. The main objective of supervised learning algorithms is to generate a robust and generalized model that can work well not only on seen data (training data) but also predict well on unseen data (test data). One of the main issues with supervised learning approaches is model overfitting. Model overfitting occurs when there is insufficient training data, or training data is too simple to cover the structural complexity of the domain being modelled. In overfitting, the final model works well on training data because the model is specialized on training data but provides significantly inaccurate predictions on test data due to the model’s lack of generalization capabilities. In this paper, we propose a novel approach to address this model overfitting that is inspired by the processes of natural immune systems. Here, we propose that the issue of overfitting can be addressed by generating more data samples by analyzing existing scarce data. The proposed approach is tested on benchmarked datasets using two different classifiers, namely, artificial neural networks and C4.5 (decision tree algorithm).