Domain Adaptation with Factorizable Joint Shift
This addresses domain adaptation for machine learning applications where biases exist in both features and labels, but it appears incremental as it builds on existing methods.
The paper tackles domain adaptation when both covariates and labels have sampling biases across domains, proposing the Factorizable Joint Shift (FJS) assumption and Joint Importance Aligning (JIA) method to improve importance estimation, with experiments on a synthetic dataset showing advantages.
Existing domain adaptation (DA) usually assumes the domain shift comes from either the covariates or the labels. However, in real-world applications, samples selected from different domains could have biases in both the covariates and the labels. In this paper, we propose a new assumption, Factorizable Joint Shift (FJS), to handle the co-existence of sampling bias in covariates and labels. Although allowing for the shift from both sides, FJS assumes the independence of the bias between the two factors. We provide theoretical and empirical understandings about when FJS degenerates to prior assumptions and when it is necessary. We further propose Joint Importance Aligning (JIA), a discriminative learning objective to obtain joint importance estimators for both supervised and unsupervised domain adaptation. Our method can be seamlessly incorporated with existing domain adaptation algorithms for better importance estimation and weighting on the training data. Experiments on a synthetic dataset demonstrate the advantage of our method.