A minimum description length technique for semi-supervised time series classification

Abstract

In recent years the plunging costs of sensors/storage have made it possible to obtain vast amounts of medical telemetry, both in clinical settings and more recently, even in patient’s own homes. However for this data to be useful, it must be annotated. This annotation, requiring the attention of medical experts is very expensive and time consuming, and remains the critical bottleneck in medical analysis. The technique of Semi-supervised learning is the obvious way to reduce the need for human labor, however, most such algorithms are designed for intrinsically discrete objects such as graphs or strings, and do not work well in this domain, which requires the ability to deal with real-valued objects arriving in a streaming fashion. In this work we make two contributions. First, we demonstrate that in many cases a surprisingly small set of human annotated examples are sufficient to perform accurate classification. Second, we devise a novel parameter-free stopping criterion for semi-supervised learning. We evaluate our work with a comprehensive set of experiments on diverse medical data sources including electrocardiograms. Our experimental results suggest that our approach can typically construct accurate classifiers even if given only a single annotated instance.