Deep Masking Generative Network For Irregularly Sampled Multivariate Time Series

Irregularly sampled multivariate time series data arise naturally in many application domain where they present a significant challenge to standard deep learning models, because they do not naturally yield a fixed-dimensional representation. In this paper, we propose a new deep learning framework named Deep Masking Generative Network from the perspective of missing data. And we introduce masking layer and unmasking layer to convert the sampled multivariate time series data into a fixed dimensional temporal representation and obtain the sampled values at any time point respectively. To process continuous time series data efficiently, we further propose continuous convolution layers with tunable kernel width. We investigate the performance of our framework on classification tasks using real datasets. Experiments show that our model performs better than a range of baseline while offering significantly faster training times.