Fair Transfer Learning Through Causally Motivated Regularization

Dylan Zapzalka

Abstract:

Transfer learning is where a source model trained on one domain is adapted for
a downstream task on another domain. Recently, it has been shown that the
unfair behaviors of the source model can persist even after it has been adapted
for a downstream task. In this work, we propose a solution to this problem by
using causally-motivated regularization schemes for creating fair source models
through using auxiliary labels. Our regularization schemes work by enforcing
independences with respect to the causal DAG. Our approach only requires having
auxiliary labels at the time of source model training and it promotes adapted
downstream models that don’t make predictions based off of sensitive attributes.
We show empirically and theoretically that source models that use our proposed
causally-motivated regularization schemes lead to fairer downstream models and
require less data to adapt to other tasks.