Domain Adaptation With Additional Features via Label-Aware and Graph-Based Fused Gromov-Wasserstein Optimal...

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Domain adaptation typically assumes shared feature spaces, challenged by new target-only features at test time.
• Existing methods often fail when target-specific features emerge post-training, violating core assumptions.
• Transductive domain adaptation requires joint inference for unlabeled target data available during training.

Purpose of the Study:

• To develop a novel domain adaptation method, LAGB-FGW, capable of handling new target-only features.
• To address the challenge of distributional shift and feature space discrepancies in domain adaptation.
• To improve label transfer accuracy in scenarios with evolving target data characteristics.

Main Methods:

• Label-Aware and Graph-Based Fused Gromov-Wasserstein Optimal Transport (LAGB-FGW) is proposed.
• The method embeds label discrepancy into the source metric and uses a K-NN graph on target features.
• It jointly solves standard Optimal Transport (OT) and Gromov-Wasserstein OT for label transfer.

Main Results:

• LAGB-FGW demonstrated consistent superior performance across four synthetic benchmarks.
• The method achieved state-of-the-art results on the HAR70+ human-activity dataset.
• Outperformance highlights the benefit of combining source label information with graph-based structural cues.

Conclusions:

• LAGB-FGW effectively adapts models when new target features are available only at test time.
• The approach successfully leverages both common and additional features for robust label transfer.
• This method offers a significant advancement for domain adaptation in dynamic environments.