A review of domain adaptation without target labels (1901.05335v2)

Abstract: Domain adaptation has become a prominent problem setting in machine learning and related fields. This review asks the question: how can a classifier learn from a source domain and generalize to a target domain? We present a categorization of approaches, divided into, what we refer to as, sample-based, feature-based and inference-based methods. Sample-based methods focus on weighting individual observations during training based on their importance to the target domain. Feature-based methods revolve around on mapping, projecting and representing features such that a source classifier performs well on the target domain and inference-based methods incorporate adaptation into the parameter estimation procedure, for instance through constraints on the optimization procedure. Additionally, we review a number of conditions that allow for formulating bounds on the cross-domain generalization error. Our categorization highlights recurring ideas and raises questions important to further research.

• The paper reviews unsupervised domain adaptation methods, categorizing them into sample-based, feature-based, and inference-based approaches to address the absence of target labels.
• It details techniques such as data and class importance-weighting, subspace mapping, and adversarial training to minimize domain discrepancies.
• The analysis underscores challenges in assumption validation and inspires future research for robust cross-domain generalization in label-scarce environments.

Review of Domain Adaptation Without Target Labels

The paper "A Review of Domain Adaptation Without Target Labels" by Wouter M. Kouw and Marco Loog provides a systematic examination of domain adaptation, where the objective is to generalize a classifier trained on a source domain to a target domain without access to labeled data from the latter. This approach is particularly critical for machine learning applications where acquiring labeled data in the target domain is challenging or expensive.

Summary of Methods

The review categorizes domain adaptation methods into three primary categories based on their methodological focus: sample-based, feature-based, and inference-based approaches.

1. Sample-Based Approaches: These strategies rely on weighting individual samples from the source domain to mimic the target distribution as closely as possible. This category includes:
   • Data Importance-Weighting: Assumes a covariate shift, where the conditional probability distribution remains consistent while the marginal distribution shifts across domains. Methods under this category focus on estimating the importance weights to emphasize samples that are more representative of the target domain.
   • Class Importance-Weighting: Assumes a prior shift, where class priors differ between domains while the class-conditional distributions remain consistent. Techniques here estimate how class distributions change between the domains.
2. Feature-Based Approaches: These focus on transforming the feature space to reduce the discrepancy between domains without altering the classifier architecture:
   • Subspace Mappings and Optimal Transport: Attempt to align source data with the target data by finding projections or transformations that map one to the other. Optimal transport aims to minimize the cost of aligning source and target distributions.
   • Domain-Invariant Spaces: Techniques to project both domains into a shared space that preserves class-discriminative features while eliminating domain-specific noise.
   • Deep Domain Adaptation: Networks trained with an adversarial loss to reduce domain discrepancies while maintaining predictive performance.
   • Correspondence Learning: Identifies and uses shared feature dependencies across domains to construct domain-agnostic features.
3. Inference-Based Approaches: These techniques embed adaptation into the parameter estimation process:
   • Algorithmic Robustness and Minimax Estimators: Focus on making models robust to distributional changes between domains, often by minimizing the worst-case risk.
   • Self-Learning: Uses iterative self-labeling where a model trained on source data is used to label and include target domain samples progressively.
   • Bayesian Methods: Utilizes the source domain to inform prior distributions in Bayesian models, leveraging prior knowledge to guide inference in the target domain.

Implications and Future Directions

The review emphasizes the importance of assumptions underlying each method, noting that these assumptions cannot be validated without access to target labels, thus highlighting the inherent uncertainty in unsupervised domain adaptation. Nonetheless, this area of research has substantial implications, particularly for fields where labeled data is sparse or costly to obtain, such as biomedical imaging or natural language processing.

Theoretical insights derive from generalization error bounds that consider differences between source and target domains, which are mainly influenced by the chosen assumption on data shifts (covariate vs. prior shift). Future research directions include developing robust hypothesis tests for validating domain adaptation assumptions and exploring causal relationships within data to better inform transformations between domains.

Conclusion

The review by Kouw and Loog provides a comprehensive overview of existing domain adaptation techniques without target labels, elucidating the intricacies of transferring knowledge across domains. As the need for cross-domain generalization continues to rise across various application domains, the insights and frameworks provided in this paper serve as a valuable resource for developing effective adaptation strategies that leverage available labeled data efficiently.

In conclusion, while the field has made significant strides, addressing the challenge of assumption validation and better understanding of causal inferences promises to enhance the reliability of domain-adaptive models in the future.