Abstract
Source-Free Domain Adaptation (SFDA) addresses the challenge of adapting a pre-trained source model to a target domain without access to original source data, a common requirement in privacy-sensitive applications. Active SFDA (ASFDA) further enhances this by allowing a small fraction of target data to be labeled. However, existing active learning methods often struggle with model miscalibration on the target domain. We propose Minimum Happy Points Learning (MHPL), a novel framework that identifies the most informative "Minimum Happy" (MH) points—samples that are source-dissimilar and reside in chaotic neighborhood environments. By leveraging local neighbor structures rather than unreliable model confidence, MHPL achieves significant performance gains across multiple benchmarks, including a 6.5% improvement on the Office-Home dataset with only a 5% labeling budget.
Introduction
In many real-world scenarios, such as healthcare or security, data privacy prevents the sharing of raw source datasets. Source-Free Domain Adaptation (SFDA) has emerged to handle these constraints, but it often faces a performance bottleneck due to the lack of direct domain discrepancy measurements. Active SFDA (ASFDA) attempts to overcome this by incorporating a human-in-the-loop to label a tiny portion (e.g., 5%) of the target data.
Standard active learning strategies typically rely on uncertainty metrics like entropy or least confidence. In the context of domain shift, these metrics are often unreliable because the source model is miscalibrated on the target domain, leading to overconfident but incorrect predictions. MHPL addresses this by focusing on the "unhappiest" points—those that the model finds most difficult based on their local feature-space environment.
Methodology: Identifying Minimum Happy Points
The core of MHPL is the selection of Minimum Happy (MH) points. These are samples that are neighbor-chaotic, diverse, and source-dissimilar. The framework consists of four primary components:
- Neighbor Environment Uncertainty (NEU): Instead of relying on the model's self-prediction, NEU evaluates a sample's neighborhood. It combines Neighbor Purity (label consistency among neighbors) and Neighbor Affinity (spatial proximity) to identify samples in chaotic regions where the model is likely struggling.
- Neighbor Diversity Relaxation (NDR): To ensure the selected active set covers the target distribution effectively, NDR prevents the selection of multiple samples from the same local neighborhood.
- One-shot Querying: MHPL utilizes the raw source model for a single round of querying. The authors find that the initial, un-adapted model serves as the best "yardstick" for identifying samples that are truly dissimilar to the source domain.
- Neighbor Focal (NF) Loss: During the adaptation phase, this loss function prioritizes learning from the labeled MH points, weighting them by their neighbor purity to ensure robust generalization.
Experimental Results
The MHPL framework was evaluated on three major domain adaptation benchmarks: Office-Home, VisDA-2017, and Office-31.
Performance Gains
MHPL consistently outperformed both state-of-the-art SFDA methods and established Active Learning strategies. On the Office-Home dataset, MHPL achieved an average accuracy of 79.3%, which is a 6.5% improvement over the SFDA method A2Net and a 2.3% lead over the previous best ASFDA method, ELPT.
Label Efficiency
A key finding of the study is the high label efficiency of MHPL. With only a 5% labeling budget, the method effectively bridges the gap between source-free adaptation and methods that require full source data access. This demonstrates that identifying "Minimum Happy" points is a superior strategy for selecting informative samples in the presence of domain shift.
Conclusion
MHPL introduces a robust approach to Active Source-Free Domain Adaptation by shifting the focus from model-centric uncertainty to data-centric neighbor structures. By identifying "Minimum Happy" points that are both source-dissimilar and neighborhood-chaotic, the framework overcomes the common pitfall of model miscalibration. Future work may involve optimizing the computational cost of neighbor searches for large-scale datasets and further exploring the theoretical properties of source-dissimilarity in active learning.
Summary updated