Time-series classification, with its vast applications in many fields, such as weather prediction, medicine, zoology, and human behavior analysis, is an active area of research. The problems in this field typically involve training a classifier on a set of time series, with each series consisting of real values and the corresponding class label, where the real values have a natural temporal ordering, such as a patient’s systolic blood pressure at the end of every minute.
This work falls within the specific subarea of never-ending learning models. It is suitable for researchers in this field (graduate students, those doing their post-docs, faculty, and so on), as well as senior practitioners in the field of time series classification due to its significant focus on experimental results.
In this category of learning models, “an agent examines an unbounded stream of data” (as opposed to a finite set, such as the last 36 months or the last 20 minutes). The best-known work in the field of never-ending learning models is by the NELL team [1] at Carnegie Mellon University, processing “millions of web pages looking for connections between the information it already knows and what it finds through its search process.” However, there are notable differences between Chen et al.’s work and the NELL team’s work. The NELL team processes discrete data and can process the data over and over again in order to create a useful ontology, while Chen et al.’s work looks at the data only once. Another relevant work is Berlin and Laerhoven’s [2], although that work requires significant human input and is not a true continuous learning system.
One of the strengths (and perhaps the weakness as well) of Chen et al.’s work is that it draws most of its conclusions from experimental results on actual data sets. A substantial part of the paper is dedicated to experimental results, and a very well-documented separate website (https://sites.google.com/site/nelframework/) supports the experimental results as well. The different experiments vary from studies of insects, to electrocardiogram (ECG) patterns, to energy disaggregation. The supporting website does an amazing job in providing the actual data sets, code examples, and practical suggestions on when to invoke queries and so on. However, on the same note, I did not find some of the assertions in Section 3.5 (“Frequent Pattern Maintenance”) to be general enough. While the authors note that some of the results were borne out of real experiments, a characterization has not been attempted on the generality of those results, and that can be a direction for future theoretical extensions of this paper.