Prototypical Partial Optimal Transport for Universal Domain Adaptation

TL;DR

This work tackles Universal Domain Adaptation (UniDA) by formulating partial distribution alignment between source and target domains using mini-batch Prototypical Partial Optimal Transport (m-PPOT). The approach builds source-class prototypes and performs partial OT to align common-class mass, while leveraging the transport plan to reweight source prototypes and target samples via reweighted cross-entropy and entropy losses to distinguish known versus unknown samples. The authors provide theoretical connections between POT, m-PPOT, and prototype distances, and demonstrate state-of-the-art performance on four UniDA benchmarks with extensive ablations showing the effectiveness of the m-PPOT loss, reweighting strategies, and contrastive pre-training. The method offers a scalable, principled partial alignment framework for UniDA with strong empirical impact on OPDA, PDA, and OSDA tasks across large and diverse datasets.

Abstract

Universal domain adaptation (UniDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain without requiring the same label sets of both domains. The existence of domain and category shift makes the task challenging and requires us to distinguish "known" samples (i.e., samples whose labels exist in both domains) and "unknown" samples (i.e., samples whose labels exist in only one domain) in both domains before reducing the domain gap. In this paper, we consider the problem from the point of view of distribution matching which we only need to align two distributions partially. A novel approach, dubbed mini-batch Prototypical Partial Optimal Transport (m-PPOT), is proposed to conduct partial distribution alignment for UniDA. In training phase, besides minimizing m-PPOT, we also leverage the transport plan of m-PPOT to reweight source prototypes and target samples, and design reweighted entropy loss and reweighted cross-entropy loss to distinguish "known" and "unknown" samples. Experiments on four benchmarks show that our method outperforms the previous state-of-the-art UniDA methods.

Figures (11)

• Figure 1: Illustration of our model. Source and target data share the same feature extractor that embeds data in feature space. PPOT is to match target features and source prototypes which are updated by the source features, and the row/column sum of transport plan is applied for reweighting. We design reweighted entropy loss to align common class features of two domains, while pushing away the unknown features.
• Figure 2: (a) Class weight $\bm{w}^s$ in Eqn. (\ref{['eq: rceloss']}) on the source domain. (b) Average weight $\bm{w}^t$ in Eqn. (\ref{['eq: peloss']}) for each class on the target domain. Task: W$\rightarrow$D on Office-31 for OPDA.
• Figure 3: Sensitivity to hyper-parameters (a) $\eta_1$, $\eta_2$ and $\eta_3$ in Eqn. (\ref{['eq: total']}), (b) $\tau_1$, $\tau_2$ in Eqn. (\ref{['eq: alpha']}) and threshold $\xi$. All results are for the OPDA setting in task C$\rightarrow$A.
• Figure 4: H-score curves of different methods with varying number of target private classes for OPDA tasks A$\rightarrow$P and P$\rightarrow$R.
• Figure A-1: Source class weight computed by the row sum of solutions of $\text{PPOT}^{\alpha}(\bm{p}, \bm{q})$ and $\text{m-PPOT}^{\alpha}(\bm{p}, \bm{q})$ for OPDA task D$\rightarrow$W.

Theorems & Definitions (11)

• Definition 1
• Proposition 1
• Theorem 1
• Proposition 1
• proof
• Theorem 1
• Lemma 2
• proof
• Lemma 3
• proof