Semi-supervised Sequence Classification through Change Point Detection

Sequential sensor data is generated in a wide variety of real-world applications. A fundamental machine learning challenge involves effective classifiers for such sequential data. While deep has led to impressive performance gains recent years within domains as speech, this relied on the availability large datasets sequences with high-quality labels. In many applications, however, associated class labels are often extremely limited, precise labelling/segmentation being too expensive perform high volume. However, amounts unlabelled may still be available. paper we propose novel framework semi-supervised contexts. an unsupervised manner, change-point detection methods can used identify instances where classes change sequence. We show that points provide examples similar/dissimilar pairs which, when coupled labels, classification setting. Pairs from and by neural network learn improved representations classification. extensive synthetic simulations learned better than those through autoencoder obtain results human activity recognition datasets.