Unsupervised Intra-Domain Adaptation for Recommendation via Uncertainty Minimization

Given labeled data on a source domain, unsupervised domain adaptation (UDA) for recommendation aims at making desirable prediction for unlabeled data on a target domain which is considerably different from the source domain. Mainstream approaches to UDA for recommendation consist in learning aligned features and minimizing transferred knowledge discrepancy between two domains. However, these methods which focus on overall domain discrepancy tend to ignore intrinsic intra-domain discrimination. Motivated by the uncertainty representation theory, we propose to contrastively minimize the intrinsic intra-domian discrepancy between two domains via uncertainty minimization oriented learning. We constrain the knowledge transfer via uncertainty minimization that hinges on uncertainty representation determined by the pivot similarity between different network branches. More specifically, the Kullback-Leibler (KL) divergence and Euclidean square distance between predicted label distributions of the main classifier and an introduced auxiliary classifier are used to characterize prediction variances. We involve the prediction variance as uncertainty representation into optimization objectives. Experimental results on datasets demonstrate that our proposed method outperforms state-of-the-art methods.