Discovering latent causal variables using a trade-off between compression and causality

Causality is a fundamental relationship in the physical world, around which almost all activities of human life revolve. Causal inference refers to the process of determining whether an event or action caused a specific outcome, which involves the evaluation of cause-and-effect relationships in data. This paper presents a new approach to discover latent causal representations of crucial variables in easy-to-obtain data. The proposed method takes a form of trade-off between compression of input data and the causality between the learnt latent variables and critical variables, thereby removing the irrelevant information contained in input data and obtaining the decoupled, strongest causal factors. By introducing variational bounds and specific configurations, the optimisation objective is relaxed to a tractable problem. The approach compacts causal discovery and inference into one model, which is flexible to downstream tasks and parsimonious in the parameters. A case study on an exhaust-emission dataset shows that the proposed method improves the predictive performance over the baseline model, which is a variational information bottleneck model with the same hyperparameters.