Causal Recurrent Variational Autoencoder for Medical Time Series Generation

We propose causal recurrent variational autoencoder (CR-VAE), a novel generative model that is able to learn Granger graph from multivariate time series x and incorporates the underlying mechanism into its data generation process. Distinct classical VAEs, our CR-VAE uses multi-head decoder, in which p-th head responsible for generating dimension of (i.e., x^p). By imposing sparsity-inducing penalty on weights (of decoder) encouraging specific sets be zero, learns sparse adjacency matrix encodes relations between all pairs variables. Thanks this matrix, decoder strictly obeys principles causality, thereby making process transparent. develop two-stage approach train overall objective. Empirically, we evaluate behavior synthetic two real-world human brain datasets involving, respectively, electroencephalography (EEG) signals functional magnetic resonance imaging (fMRI) data. Our consistently outperforms state-of-the-art models both qualitatively quantitatively. Moreover, it also discovers faithful with similar or improved accuracy over existing causality-based inference methods. Code publicly available at https://github.com/hongmingli1995/CR-VAE.