Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Heterogeneous Domain Adaptation with Positive and Unlabeled Data

Junki Mori, Ryo Furukawa, Isamu Teranishi, Jun Sakuma

TL;DR

This work defines PU-HUDA, a challenging setting where the source domain provides only positive labels and the target domain is unlabeled under heterogeneous feature spaces. It introduces Predictive Adversarial Domain Adaptation (PADA), a unified adversarial framework that learns a feature transformer, a classifier, and a discriminator to align the source with likely-positive target data via KL-divergence objectives, while soft-labeling guides training. A soft-labeling mechanism leveraging a base PU learner on common features accelerates and stabilizes learning, enabling iterative refinement. Across Movielens-Netflix, 20-Newsgroups, and credit-card datasets, PADA and its soft-labeling variant outperform baseline PU-HUDA and PU methods, demonstrating effective transfer under label scarcity and feature heterogeneity.

Abstract

Heterogeneous unsupervised domain adaptation (HUDA) is the most challenging domain adaptation setting where the feature spaces of source and target domains are heterogeneous, and the target domain has only unlabeled data. Existing HUDA methods assume that both positive and negative examples are available in the source domain, which may not be satisfied in some real applications. This paper addresses a new challenging setting called positive and unlabeled heterogeneous unsupervised domain adaptation (PU-HUDA), a HUDA setting where the source domain only has positives. PU-HUDA can also be viewed as an extension of PU learning where the positive and unlabeled examples are sampled from different domains. A naive combination of existing HUDA and PU learning methods is ineffective in PU-HUDA due to the gap in label distribution between the source and target domains. To overcome this issue, we propose a novel method, predictive adversarial domain adaptation (PADA), which can predict likely positive examples from the unlabeled target data and simultaneously align the feature spaces to reduce the distribution divergence between the whole source data and the likely positive target data. PADA achieves this by a unified adversarial training framework for learning a classifier to predict positive examples and a feature transformer to transform the target feature space to that of the source. Specifically, they are both trained to fool a common discriminator that determines whether the likely positive examples are from the target or source domain. We experimentally show that PADA outperforms several baseline methods, such as the naive combination of HUDA and PU learning.

Heterogeneous Domain Adaptation with Positive and Unlabeled Data

TL;DR

This work defines PU-HUDA, a challenging setting where the source domain provides only positive labels and the target domain is unlabeled under heterogeneous feature spaces. It introduces Predictive Adversarial Domain Adaptation (PADA), a unified adversarial framework that learns a feature transformer, a classifier, and a discriminator to align the source with likely-positive target data via KL-divergence objectives, while soft-labeling guides training. A soft-labeling mechanism leveraging a base PU learner on common features accelerates and stabilizes learning, enabling iterative refinement. Across Movielens-Netflix, 20-Newsgroups, and credit-card datasets, PADA and its soft-labeling variant outperform baseline PU-HUDA and PU methods, demonstrating effective transfer under label scarcity and feature heterogeneity.

Abstract

Heterogeneous unsupervised domain adaptation (HUDA) is the most challenging domain adaptation setting where the feature spaces of source and target domains are heterogeneous, and the target domain has only unlabeled data. Existing HUDA methods assume that both positive and negative examples are available in the source domain, which may not be satisfied in some real applications. This paper addresses a new challenging setting called positive and unlabeled heterogeneous unsupervised domain adaptation (PU-HUDA), a HUDA setting where the source domain only has positives. PU-HUDA can also be viewed as an extension of PU learning where the positive and unlabeled examples are sampled from different domains. A naive combination of existing HUDA and PU learning methods is ineffective in PU-HUDA due to the gap in label distribution between the source and target domains. To overcome this issue, we propose a novel method, predictive adversarial domain adaptation (PADA), which can predict likely positive examples from the unlabeled target data and simultaneously align the feature spaces to reduce the distribution divergence between the whole source data and the likely positive target data. PADA achieves this by a unified adversarial training framework for learning a classifier to predict positive examples and a feature transformer to transform the target feature space to that of the source. Specifically, they are both trained to fool a common discriminator that determines whether the likely positive examples are from the target or source domain. We experimentally show that PADA outperforms several baseline methods, such as the naive combination of HUDA and PU learning.
Paper Structure (23 sections, 4 equations, 4 figures, 6 tables, 1 algorithm)

This paper contains 23 sections, 4 equations, 4 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Heterogeneous Unsupervised Domain Adaptation
  4. PU learning
  5. Domain Adaptation in PU learning setting
  6. Positive and Unlabeled Heterogeneous Domain Adaptation
  7. Problem Definition
  8. Difficulty of PU-HUDA
  9. Proposed Approach
  10. Basic Idea
  11. Naive Optimization Approach
  12. Predictive Adversarial Domain Adaptation
  13. Soft labeling mechanism to improve PADA
  14. Experiments
  15. Experimental Setup
  16. ...and 8 more sections

Figures (4)

  • Figure 1: View of the data structure in two HUDA settings, (a) conventional one and (b) new one where the source domain has only positive data. Here, both of them assume that the two domains share some features.
  • Figure 2: Comparison of adaptation mechanisms in existing HUDA methods (top) and our proposed PU-HUDA method (bottom).
  • Figure 3: Overview of the proposed method, PADA using soft-labeling mechanism.
  • Figure 4: Effect of the number of selected common features for Movielens-Netflix