ENSEMBLE-BASED OUTLIER DETECTION

Outlier detection is the process of detecting observations that do not conform to the norm of the dataset. Outlier detection is an unsupervised and thus ill-posed problem. This makes the outlier detection task particularly challenging. Ensembles have the potential to address the ill-posed nature of outlier detection. Ensemble techniques have proven to be effective for classification and clustering task yet anomaly ensembles have only recently been studied. In this dissertation, I present strategies to develop robust and accurate ensemble methods including neural network-based approaches for detecting outliers. Specifically, I introduce (i) Soul, a framework for building selective outlier ensembles that has shown potential for minimizing the negative impact of poor components in an ensemble, (ii) BAE, an unsupervised boosting-based ensemble approach that is proposed to overcome limitations of using autoencoders in outlier detection, and (iii) \ExVAE, an exemplar-based variational autoencoder with a novel application in the field of outlier detection, tailor-made to tackle the challenges posed in this domain.

Author

• Sarvari, Hamed

• Deep learning
• Ensemble methods
• Machine learning
• Outlier detection

Identifier

hdl:1920/13775

OAI identifier oai:identifier

oai:MARS:1920/13775