Many time series data mining algorithms are designed to be intuitive for domain experts, often including case studies and benchmark results. However, it is challenging to create truly domain-agnostic algorithms due to unique domain-specific invariances. In these works, we prioritize and leverage user input and interpretability over complexity. We empirically demonstrate that many time series data mining tasks are still a searchable space, allowing for interpretable methodologies. We adopt a ``Soup-to-Nuts" design, envisioning how the method fits into a user's problem, from accepting their real-valued time series dataset to providing a direct answer to the data mining task. With these three characteristics in mind, we propose methods that are user-friendly, not requiring the user to have extensive experience with time series data mining methods and find the results of our methods actionable. In the first work, we target a problem setting of long time series. Many time series datasets have been cleaned and preprocessed to be of uniform length and have guarantees that the time series contains some semantic data the dataset advertises. We relax these two constraints, considering signals without any assumptions about their contents. In the second work, we propose a time series anomaly explanation method and format. With the increasing relevance of time series anomaly detection, we formulate an intuitive and actionable anomaly explanation method, addressing the often complicated and unclear actionability of existing explanations. In the third work, we propose a method for time series clustering, leveraging feature sets for interpretability and user input for customizability, guiding our methodology to make better-suited and more intelligible decisions for users based on their specific needs.

This dissertation is motivated from enabling various tasks in large scale data mining of time series to produce more accurate, reproducible results and tailored to user’s specific need when that is favored. To that end, we have explored and contributed to the literature in three parts; each touches an active area of research and unifies under a common theme, reducing errors in time series data mining by learning constraints on model’s flexibility.

The first body of work concerns Dynamic Time Warping (DTW), a highly competitive distance measure for most time series data mining problems. Obtaining the best performance from DTW requires setting its only parameter, the maximum amount of warping. This parameter gives DTW the flexibility to deal with data that can be locally out of phase, however the DTW algorithm sometimes exploits this flexibility to give pathological and unwanted results. We demonstrate the importance of setting DTW’s warping window width correctly, to constrain this flexibility, and we propose novel methods to learn this parameter in both supervised and unsupervised settings.

The second body of work concerns time series motif discovery, perhaps the most used primitive for time series data mining. We point out that the current definitions of motif discovery are limited and can create a mismatch between the user’s intent/expectations, and the motif discovery search outcomes. We explain the reasons behind these issues and introduce a novel and general framework to address them.

The last body of work concerns making more time series data sets and baseline results publicly available for gauging progress and comparison of rival approaches in spirit of reproducible research. We work on expanding the UCR Time Series Archive, an important resource in the time series data mining community, from 85 data sets since the last Fall 2015 release to 128 data sets in Fall 2018. Creating benchmark results for this archive required 61,041,100,000,000 DTW comparisons, greatly more than the number of DTW comparisons that have appeared in all research papers combined. Beyond expanding this valuable resource, we offer pragmatic advice to anyone who may wish to evaluate a new algorithm on the archive.

Interdisciplinary research in computer science requires the development of computational techniques for practical application in different domains. This usually requires careful integration of different areas of technical expertise. This dissertation presents image and time series analysis algorithms, with practical interdisciplinary applications to develop-mental biology, historical manuscript processing, and data stream processing. Inspired by the NSF IGERT program, this dissertation presents algorithms for analysis of growth dy-namics at the shoot apex of Arabidopsis thaliana. A robust understanding of the causal relationship between gene expression, cell behaviors, and organ growth requires the de-velopment of computational techniques for quantitative analysis of real-time, live-cell meristem growth data. This requires the development/application of image analysis tools and novel time series alignment algorithms. Image analysis is necessary for the computa-tion of growth features, but this leads to a time series of unsynchronized growth data, which requires a robust alignment method. Towards this end, we present two time series alignment algorithms. This dissertation further considers image mining in historical doc-ument processing. An application of the Minimum Description Length principle (MDL) to develop a symbols clustering algorithm is presented. The developed algorithm pro-duced one of the first practical applications of MDL to real-world, real-valued data such as images. Moreover, we introduce a novel premise that a clustering algorithm should have the freedom to ignore some data. Extensive empirical results show that the MDL-based algorithm outperforms the popular K-Means clustering algorithm, given the same input data, distance measure, and the correct value of K in K-means. The new algorithm could have significant impact, as clustering is a critical subroutine in almost all historical document processing systems. Finally, we present an algorithm for detecting rare and ap-proximately repeating sequences in unbounded real-valued data streams, given limited space. This algorithm employs the novel integration of SAX time series representation with a Bloom filter to develop a robust cache maintenance policy that allows us to over-come known challenges to a previously unsolved frequent pattern mining problem. Our contribution lies in the fact that we solve this problem for real-valued data, whereas only the discrete-valued case has been considered in the literature.