PAC-Bayesian Domain Adaptation Bounds for Multi-view learning
Mehdi Hennequin, Khalid Benabdeslem, Haytham Elghazel
TL;DR
The paper addresses unsupervised domain adaptation in a multi-view setting by embedding MV learning into the PAC-Bayesian framework. It introduces a novel multi-view domain distance and develops a general MV-DA PAC-Bayes theorem, enabling bounds on the target risk that incorporate both distributional disagreement across views and cross-view joint-errors. It then specializes these MV bounds to classical PAC-Bayes formulations (McAllester, Catoni, etc.) and extends the theory to a two-level MV framework, yielding a PAC-Bayesian MV-DA bound with corollaries. The work provides a rigorous theoretical foundation for MVDA generalization, highlighting the role of view-specific priors/posteriors and the benefit of averaging over learning samples, and sets the stage for future MV-based domain adaptation algorithms.
Abstract
This paper presents a series of new results for domain adaptation in the multi-view learning setting. The incorporation of multiple views in the domain adaptation was paid little attention in the previous studies. In this way, we propose an analysis of generalization bounds with Pac-Bayesian theory to consolidate the two paradigms, which are currently treated separately. Firstly, building on previous work by Germain et al., we adapt the distance between distribution proposed by Germain et al. for domain adaptation with the concept of multi-view learning. Thus, we introduce a novel distance that is tailored for the multi-view domain adaptation setting. Then, we give Pac-Bayesian bounds for estimating the introduced divergence. Finally, we compare the different new bounds with the previous studies.