Rbm-od: a restricted Boltzmann machine framework for outlier detection

Abstract

This thesis explores the use of Restricted Boltzmann Machines (RBMs), a class of unsupervised generative neural networks, for detecting outliers through data generation and representation-based comparison. Outlier detection (OD) is a critical task in domains where rare or anomalous patterns may indicate errors, fraud, or unexpected behaviour in data. The primary contribution of this work is a unified framework for RBM-based outlier detection that emphasizes data generation as a detection strategy. We explore multiple model variants, including single RBMs, ensembles of RBMs, stacked RBMs, and ensembles of stacked RBMs, each offering distinct advantages in representing complex data patterns. By generating synthetic samples from trained RBMs and comparing them to input data, the approach enables unsupervised detection of unusual or unexpected instances. This gener ative perspective distinguishes RBM-OD from traditional methods and provides a flexible foundation for future extensions.