Outlier Detection Using Autoencoder Ensembles: A Robust Unsupervised Approach

Outlier analysis finds its applicability in multiple domains like finance, health, and manufacturing. As data keeps on increasing, it gets more challenging to have labels on them. Thus, unsupervised outlier analysis has started gaining center stage. The current paper makes use of an ensemble of autoencoders for outlier detection in a robust way. Autoencoders are suitable for dimensionality reduction, and they allow mapping of higher-order features to lower-order where they can be segregated. Experiments achieve state of the art results that are comparable with other neural network-based approaches on varied datasets. Ensemble framework provides robustness as it is able to deal with different domains and takes care of the inherent problem of over-fitting that comes with any neural network.