End-to-end anomaly detection in stream data

Nowadays, huge volumes of data are generated with increasing velocity through various systems, applications, and activities. This increases the demand for stream and time series analysis to react to changing conditions in real-time for enhanced efficiency and quality of service delivery as well as upgraded safety and security in private and public sectors. Despite its very rich history, time series anomaly detection is still one of the vital topics in machine learning research and is receiving increasing attention. Identifying hidden patterns and selecting an appropriate model that fits the observed data well and also carries over to unobserved data is not a trivial task. Due to the increasing diversity of data sources and associated stochastic processes, this pivotal data analysis topic is loaded with various challenges like complex latent patterns, concept drift, and overfitting that may mislead the model and cause a high false alarm rate. Handling these challenges leads the advanced anomaly detection methods to develop sophisticated decision logic, which turns them into mysterious and inexplicable black-boxes. Contrary to this trend, end-users expect transparency and verifiability to trust a model and the outcomes it produces. Also, pointing the users to the most anomalous/malicious areas of time series and causal features could save them time, energy, and money. For the mentioned reasons, this thesis is addressing the crucial challenges in an end-to-end pipeline of stream-based anomaly detection through the three essential phases of behavior prediction, inference, and interpretation. The first step is focused on devising a time series model that leads to high average accuracy as well as small error deviation. On this basis, we propose higher-quality anomaly detection and scoring techniques that utilize the related contexts to reclassify the observations and post-pruning the unjustified events. Last but not least, we make the predictive process transparent and verifiable by providing meaningful reasoning behind its generated results based on the understandable concepts by a human. The provided insight can pinpoint the anomalous regions of time series and explain why the current status of a system has been flagged as anomalous. Stream-based anomaly detection research is a principal area of innovation to support our economy, security, and even the safety and health of societies worldwide. We believe our proposed analysis techniques can contribute to building a situational awareness platform and open new perspectives in a variety of domains like cybersecurity, and health.