Title: A twin neural model for causal inference: Applications in python
Authors: Mouloud Belbahri - University of Montreal (Canada) [presenting]
Olivier Gandouet - TD Insurance (Canada)
Abstract: The focus is on the prediction of heterogeneous treatment effects in the randomized case of causal inference. We developed a solution for a specific twin neural network architecture allowing joint optimization of counterfactual marginal probabilities. We show that this model is a generalization of the logistic interaction model. We train our models with a new loss function, defined by taking advantage of a link with the Bayesian interpretation of relative risk. We modify the stochastic gradient descent algorithm to allow sparse structured solutions. This helps training to a great extent. We show that our method is competitive with the state of the art on real data from large-scale marketing campaigns.