t-SNE based feature extraction technique for multi-layer perceptron neural network classifier

In machine learning, neural network classifiers often perform well with the training samples but shows poor performance with the test samples. This scenario is called overfitting. Overfitting reduces the generalization capability of the classifier. Hence, the objective in this article is to model a classifier with better generalization. Generalization of the classifiers can be archived by reducing the dimensionality of the input space of the data set. For dimensionality reduction, t-distributed stochastic neighbor embedding (t-SNE), a machine learning algorithm has been used in this article. Initially, a two-dimensional t-SNE network is trained to successfully attain the global geometry of the data set. Here, t-SNE algorithm models each high-dimensional original input samples to a low-dimensional (usually two-dimensional) new samples by constructing probability distribution over the pairs in such a way that similar objects are modeled by nearby instances and dissimilar objects are modeled by distant instances. It tries to minimize the Kullback Leibler divergence (KL divergence) between the original high dimensional data and low dimensional projected data. Later, the projected low dimensional data has been used by multi-layer perceptron (MLP) for the classification. The complete procedure constructs a new classifier based on t-SNE and MLP. Eight standard classification data sets have been used to compare the proposed classifier with standard MLP classifier. Comparative results exhibit the supremacy of the proposed classifier. Wilcoxon signed rank test also exhibit that proposed model based on t-SNE used for revision of feature representation revamp the execution of the classifiers.