Introduction
Traditional Domain Adaptation (DA) assumes that the source domain (labeled) and the target domain (unlabeled) share an identical set of classes. However, real-world scenarios often involve "incomplete" target data where entire categories present in the source are missing from the target. This paper introduces Partially Zero-shot Domain Adaptation (PZDA) to address this gap. The goal is to train a model that accurately classifies all categories in the target domain, including those it never encountered there during training.
The Challenge of Missing Classes
In a typical DA setup, such as a surveillance system trained on synthetic "walking" and "fighting" data, the target real-world footage might only contain "walking" because "fighting" is rare. Standard methods like the Domain Adversarial Neural Network (DANN) fail here because they attempt to force-match the source's "fighting" distribution to whatever is available in the target (e.g., "walking"). This "hopeless matching" leads to feature misalignment and degrades classifier performance across the board.
Proposed Method: Instance-Weighted Adversarial Training
To solve the PZDA problem, the authors propose a novel framework that uses instance weighting to estimate the missing target data distributions. The architecture builds upon DANN, which consists of a feature extractor $G_f$, a label classifier $G_y$, and a domain discriminator $G_d$.
Core Innovations
- Loss Estimation for Unseen Data: Since target data for missing classes is unavailable, the model uses its own class posterior predictions ($p(y|x)$) to weight source samples. This allows the model to identify source samples that likely resemble the missing target classes and use them to "fill in the blanks" for the domain discrimination loss.
- Dual Weighting Mechanism: Instance weights are applied to both the classification loss ($L_y^w$) and the domain discrimination loss ($L_d^w$). This ensures the classifier does not over-fit to only the classes present in the target while preventing the discriminator from being misled by the missing categories.
- Stability Controls: To manage the potential instability of probability ratios, the authors implement weight clipping and a confidence coefficient ($eta$) to limit the influence of estimated data.
Experimental Results
The researchers evaluated PZDA on standard benchmarks modified to simulate missing classes:
- Digit Classification (MNIST to MNIST-M): When 7 out of 10 classes were removed from the target, the proposed method achieved 81.0% accuracy, vastly outperforming standard DANN, which dropped to 32.6%.
- Object Recognition (Office-31): In the Amazon-to-Webcam transition with 20 of 31 classes missing, the method maintained 63.5% accuracy, doubling the performance of the state-of-the-art Partial DA method (PADA), which scored 31.6%.
Conclusion
The PZDA framework demonstrates that by actively estimating missing data distributions through self-training and instance weighting, models can remain robust even when target domains are significantly incomplete. While the method currently assumes similar class priors between domains, it provides a vital foundation for applying domain adaptation to rare event detection and other real-world "missing class" scenarios.
Summary updated