Extended Representation Learning Based Neural Network Model for Outlier Detection

Abstract

Outlier detection problems have drawn much attention in recent times for their variety of applications. An outlier is a data point that is different from the rest of the data and can be detected based on some measure. In recent years, Artificial Neural Networks (ANN) have been used extensively for finding outliers more efficiently. This method is highly competitive with other methods currently in use such as similarity searches, density-based approaches, clustering, distance-based approaches, linear methods, etc. In this paper, we have proposed an extended representation learning based neural network. This model follows a symmetric structure like an autoencoder where the dimensions of the data are initially increased from their original dimensions and then reduced. Root mean square error is used to compute the outlier score. Reconstructed error is calculated and analyzed to detect the possible outliers. The experimental findings are documented by applying it to two distinct datasets. The performance of the proposed model is compared to several state-of-art approaches such as Rand Net, Hawkins, LOF, HiCS, and Spectral. Numerical results show that the proposed method outperforms all of these methods in terms of 5 validation scores, Accuracy (AC), Precision (P), Recall, F1 Score, AUC score.