Exploring Causality Aware Data Synthesis

Data synthesis is an important need in many real-life applications. However, generating data with preserving causal relations is a challenge. Causal relation preservation is an utmost need to boost the explainability of ML-DL based applications. Some of the challenges of preserving causality are unobserved causes in the input observations and unknown or partially known causal structures of the input data. In this work, we explore the Deep Neural Network based generative model like Variational Auto-encoder in the context of preserving and adapting causality during data generation. We emphasize the most practical scenario for data generation where there is no prior knowledge of causal structure during data generation. We have considered both observed and unobserved common causes. We have performed experimental analysis to evaluate causality preservation with increasing sample size by using our own synthetically generated data using the Structural Causal Model (SCM) and real world data like Abalone[23] from the UCI machine learning data repository.