Online Domain Indexing

ICLR 2026 Conference Submission14327 Authors

Keywords: Domain Adaptation, Probabilistic methods

TL;DR: We propose an online domain adaptation method for imbalanced data.

Abstract: Domain adaptation (DA) in real-world applications often unfolds in an online fashion, where data arrives sequentially with limited domain access and imbalanced sampling across domains. For example, in personalized ads prediction, users from different demographic groups (e.g., countries or age cohorts) correspond to distinct domains with highly skewed data availability, and user interests evolve over time. Recent work has explored domain indices to capture latent inter-domain relationships and improve adaptation (Wang et al., 2020, Xu et al., 2023). However, existing methods such as Variational Domain Index (VDI) (Xu et al., 2023) assume full domain observability and balanced mini-batches, limiting their applicability to real-world scenarios with online domain shift and data imbalance. To address these challenges, we propose Online Domain Indexing (ODI), the first continual domain indexing and adaptation framework designed for partial domain access and inter-domain sample imbalance. Starting from a base model pretrained on historical source and target domains, ODI incrementally updates domain indices over time using a smoothed reweighting kernel and a replay buffer to ensure stable adaptation. Experiments on both synthetic and real-world datasets demonstrate that ODI consistently outperforms state-of-the-art baselines in long-term accuracy under dynamic and resource-constrained conditions.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 14327