DeconICA: Deconfounding the Dataset Bias for Domain Generalization
Abstract: Domain generalization provides a research spot for enhancing the generalization capability of the machine learning model. We focus on a causal perspective for the domain generalization task. In causal theory, a confounder is a factor that affects both the cause and the effect. The confounder is often hidden, which causes problems in correctly performing the intervention. The Deconfounder approach indicates that a factorized multiple causes could be considered a substitute confounder. We choose a non-linear ICA method to factorize the data features to represent the confounder. The confounder is considered to represent the background, and domain biases. Empirical results on text and image classification domain generalization validate the proposed methods.
Paper Type: Long
Research Area: Machine Learning for NLP
Research Area Keywords: Generalization, Causality, Adversarial training
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 4608
Loading