Kunchangtai Liang

It is gaining growing attention to generate explications of machine learning algorithms that satisfy the interests of human stakeholders. Research in this area has proposed to approach it with structural causal models (SCM) that answer the question, "What action could be taken to overturn an undesirable result?" However, existing literature often makes some unrealistic assumptions, such as causal sufficiency (the absence of hidden confounders), which restricts the applications in the real world. This PhD project aims to tackle this problem by designing causal generative models with limited assumptions over the underlying SCM and the causal graph. We will extend current causal models to hidden confounders while accounting for imperfect knowledge of the world by probabilistic modelling.