Criterion Optimization-Based Unsupervised Domain Adaptation

This chapter focuses on criterion optimization-based UDA techniques. It first provides background on how most existing methods align source and target distributions by optimizing a divergence criterion like MMD. Two issues are then identified – lack of progress on new criteria and inability of statistical alignment methods to uncover causal domain-invariant factors. To address the first issue, this chapter proposes a novel divergence criterion called maximum density divergence (MDD) which jointly minimizes inter-domain divergence while maximizing intra-class density. Experiments on various benchmarks validate the effectiveness. Regarding the second issue, this chapter highlights the importance of causal features for valid adaptation. We introduce a method called joint causality-invariant feature learning (JCFL) which leverages a Hilbert-Schmidt independence criterion to identify causal factors. Extensive experiments demonstrate that JCFL consistently improves state-of-the-art methods.