Representation Learning for Anomaly Detection: From the Aspects of Data Views and Optimization

Publication Type:
Issue Date:
Full metadata record
Anomaly detection is a challenging task in realistic applications. However, there exist major issues for anomaly detection: (1) While the traditional problem setting focuses on data of single view, we have to deal with data of multiple views in many practical scenarios. Current methods often rely on assumptions on data distribution, which limits their flexibility and application. There is a lack of more effective and flexible methods for both semi-supervised and unsupervised multi-view anomaly detection. (2) Deep neural networks (DNN) have been widely applied for detecting anomalies. However, the mainstream optimization and learning strategies in DNN makes it easily fit both normal and anomalous data, resulting in an unsatisfactory performance and less reliability. There is a need for more advanced and reliable optimization framework for anomaly detection. In this thesis, we propose innovative deep representation learning models to tackle anomaly detection problems from aspects of multi-view data and model optimization. We first introduce related work and literature review. In following chapters, we investigate semi-supervised multi-view anomaly detection via variational generative model, unsupervised multi-view anomaly detection by exploring latent spaces, and advanced optimization strategy for general unsupervised anomaly detection respectively.
Please use this identifier to cite or link to this item: