PrimeNet: Pre-training for Irregular Multivariate Time Series

Real-world applications often involve irregular time series, for which the intervals between successive observations are non-uniform. Irregularity across multiple features in a multi-variate series further results different subset of at any given (i.e., asynchronicity). Existing pre-training schemes time-series, however, assume regularity and make no special treatment irregularity. We argue that such irregularity offers insight about domain property data—for example, frequency hospital visits may signal patient health condition—that can guide representation learning. In this work, we propose PrimeNet to learn self-supervised multivariate time-series. Specifically, design sensitive contrastive learning data reconstruction task pre-train model. Irregular time-series exhibits considerable variations sampling density over time. Hence, our triplet generation strategy follows original points, preserving its native Moreover, variation makes difficult regions. Therefore, masking technique always masks constant duration accommodate regions density. with these tasks using unlabeled build pre-trained model fine-tune on downstream limited labeled data, contrast existing fully supervised approach requiring large amounts data. Experiment show significantly outperforms state-of-the-art methods naturally asynchronous from Healthcare IoT several tasks, including classification, interpolation, regression.