Timezone: »

Matching Learned Causal Effects of Neural Networks with Domain Priors
Sai Srinivas Kancheti · Gowtham Reddy Abbavaram · Vineeth N Balasubramanian · Amit Sharma

Tue Jul 19 03:30 PM -- 05:30 PM (PDT) @ Hall E #616

A trained neural network can be interpreted as a structural causal model (SCM) that provides the effect of changing input variables on the model's output. However, if training data contains both causal and correlational relationships, a model that optimizes prediction accuracy may not necessarily learn the true causal relationships between input and output variables. On the other hand, expert users often have prior knowledge of the causal relationship between certain input variables and output from domain knowledge. Therefore, we propose a regularization method that aligns the learned causal effects of a neural network with domain priors, including both direct and total causal effects. We show that this approach can generalize to different kinds of domain priors, including monotonicity of causal effect of an input variable on output or zero causal effect of a variable on output for purposes of fairness. Our experiments on twelve benchmark datasets show its utility in regularizing a neural network model to maintain desired causal effects, without compromising on accuracy. Importantly, we also show that a model thus trained is robust and gets improved accuracy on noisy inputs.

Author Information

Sai Srinivas Kancheti (Indian Institute of Technology Hyderabad)

I am a second year PhD student at IITH supervised by Dr. Vineeth N Balasubramanian. I mainly work on understanding how consequential decision making systems can offer recourse to adversely affected individuals -- formally known as Algorithmic Recourse. I am also broadly interested in interpretable machine learning, causality in machine learning and decision making under strategic behaviour.

Gowtham Reddy Abbavaram (Indian Institute of Technology Hyderabad)
Vineeth N Balasubramanian (Indian Institute of Technology, Hyderabad)
Amit Sharma (Microsoft Research)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors