A Dynamic Self-Structuring Neural Network

Creating a neural network based classification model is commonly accomplished using the trial and error technique. However, the trial and error structuring method have several difficulties such as time and availability of experts. In this article, an algorithm that simplifies structuring neural network classification models has been proposed. The algorithm aims at creating a large enough structure to learn models from the training dataset that can be generalised well on the testing dataset. Our algorithm dynamically tunes the structure parameters during the training phase aiming to derive accurate non-overfitting classifiers. The proposed algorithm has been applied to phishing websites classification problem and it shows competitive results with respect to various evaluation measures such as Harmonic Mean (F1-score), precision, accuracy, etc.