Disentangled representation learning with causal effect transmission in variational autoencoder

Disentangled Representation Learning in variational autoencoder (VAE) has emerged as a strategy to identify and disentangle underlying factors from observable data to improve recognition capabilities such as images, speeches, and biological signals. Existing disentanglement methods are mostly based on the prior assumption that latent variables are mutually independent, which is inconsistent with reality and fails to transmit causal effects among causal nodes. To address the above issues, we introduce a novel disentanglement representation learning model with causal effect transmission, named DRL${}_{\mathrm{CET}}$. The main ideas of DRL${}_{\mathrm{CET}}$ involve (1) mapping encoded latent exogenous variables to causal variables and updating the causal structure by a constructed nonlinear/linear structural causal model (SCM), (2) designing hierarchical feature loss from discriminator to replace pixel-level loss in variational autoencoder for efficiently extracting causal features, and (3) aggregating causal information from adjacent nodes by a graph attention network (GAT) with intervention for transmitting causal effects. Extensive theoretical analyses and empirical studies on synthetic and real datasets demonstrate the effectiveness, viability, and superiority of our DRL${}_{\mathrm{CET}}$ over the state-of-the-arts. Our code is publicly available at https://github.com/youdianlong/DRLCET.git.