Distributed Contextual Anomaly Detection from Big Event Streams

The age of big digital data is emerged and the size of generating data is rapidly increasing in a millisecond through the Internet of Things (IoT) and Internet of Everything (IoE) objects. Specifically, most of today’s available data are generated in a form of streams through different applications including sensor networks, bioinformatics, smart airport, smart highway traffic, smart home applications, e-commerce online shopping, and social media streams. In this context, processing and mining such high volume of data stream becomes one of the research priority concern and challenging tasks. On the one hand, processing high volumes of streaming data with low-latency response is a critical concern in most of the real-time application before the important information can be missed or disregarded. On the other hand, detecting events from data stream is becoming a new research challenging task since the existing traditional anomaly detection method is mainly focusing on; a) limited size of data, b) centralised detection with limited computing resource, and c) specific anomaly detection types of either point or collective rather than the Contextual behaviour of the data. Thus, detecting Contextual events from high sequence volume of data stream is one of the research concerns to be addressed in this thesis.

As the size of IoT data stream is scaled up to a high volume, it is impractical to propose existing processing data structure and anomaly detection method. This is due to the space, time and the complexity of the existing data processing model and learning algorithms. In this thesis, a novel distributed anomaly detection method and algorithm is proposed to detect Contextual behaviours from the sequence of bounded streams. Capturing event streams and partitioning them over several windows to control the high rate of event streams mainly base on, the proposed solution firstly. Secondly, by proposing a parallel and distributed algorithm to detect Contextual anomalous event. The experimental results are evaluated based on the algorithm’s performances, processing low-latency response, and detecting Contextual anomalous behaviour accuracy rate from the event streams. Finally, to address scalability concerned of the Contextual events, appropriate computational metrics are proposed to measure and evaluate the processing latency of distributed method. The achieved result is evidenced distributed detection is effective in terms of learning from high volumes of streams in real-time.