A Minimum Description Length Technique for Semi-Supervised Time Series Classification

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