UC Riverside

Time series data is a fundamental datatype in data mining, with its understanding deepened over the past decade through analyzing subsequences based on their similarity or dissimilarity to other data. Two known concepts in this realm are motifs, which are closely resembling subsequences, and discords, which are subsequences distant from their nearest neighbors. The unique characteristic of time series data is its amenability to real-time interventions besides offline analysis. For instance, monitoring an industrial process in real-time can lead to immediate actions to prevent imminent failures, based on predictive algorithms. Various tools are available to monitor and analyze time series data for known patterns, evolving behaviors, or unexpected anomalies which are crucial for pattern matching and anomaly detection.In this dissertation, a novel analytical tool called the Contrast Profile is introduced. This tool facilitates a balanced examination of a subsequence's similarity and dissimilarity to certain data, and is demonstrated to have numerous applications including anomaly detection, data exploration, and preprocessing unstructured data for classification. The Contrast Profile enables end-to-end classification in large datasets and helps in unveiling subtle patterns that might otherwise go unnoticed. The application of the Contrast Profile is showcased across a variety of domains including seismology, animal behavior, and cardiology, demonstrating its versatility and utility in different fields. Moreover, a new primitive termed Novelets is introduced to monitor emerging behaviors in time series data. Novelets, unlike anomalies, chains, or motifs, initially appear as anomalies but are later identified as motifs. They have a broad relevance across multiple fields like science, medicine, and industry, aiding in prognostics and abnormal behavior detection. The concept of Novelets is demonstrated through various challenging domains, underlining its potential in enhancing the understanding and monitoring of emerging behaviors in time series data. Through the Contrast Profile and Novelets, this dissertation aims to significantly contribute to the field of time series data analysis, paving the way for more insightful data exploration and real-time interventions across a myriad of disciplines.