Provides in implementation of the new methods in "Invariant Models for Causal Transfer Learning". This code is in a preliminary state, please do not distribute.

How to use the code

subset_search.py consists of two functions:

subset: given training data from several tasks, returns the estimated invariant subset using Algorithm 1. subset_greedy: given training data from several tasks, returns the estimated invariant subset using Algorithm 2.

How to run a simple example

python simple_example.py

This implements a simple example using both functions. A training set with three predictors is generated, where x_1 and x_3 are causal of the target y, and x_2 is an effect of y. In each task, the coefficient used to generate x_2 from y is sampled uniformly.

How to reproduce figures in paper

Run the scripts in experiment_scripts.