Domain Adaptation for Learning from Label Proportions Using Self-Training

Ehsan Mohammady Ardehaly, Aron CulottaIJCAI 2016

Abstract

Learning from Label Proportions (LLP) is a paradigm where training data consists of groups (bags) of instances with known class distributions rather than individual labels. This paper addresses the challenge of domain adaptation in LLP, where a model trained on source domain bags must generalize to a target domain without access to target labels or bag proportions. We propose an iterative self-training approach that leverages Latent Dirichlet Allocation (LDA) to identify salient target features and construct "pseudo-bags." Our method, particularly when used in an ensemble, achieves an average 11% absolute improvement in classification accuracy across five diverse text datasets, demonstrating robustness to domain shifts and noisy pseudo-labels.

Introduction: The Challenge of LLP Domain Adaptation

In many real-world scenarios, individual labels are unavailable due to privacy concerns or the nature of data collection (e.g., census data, neighborhood-level voting statistics). Learning from Label Proportions (LLP) allows models to learn from these aggregate "bags." However, a significant bottleneck arises when the source domain (where bag proportions are known) differs from the target domain (where the model is deployed).

Standard domain adaptation techniques typically assume access to individual labels in the source domain. In the LLP setting, we face a "double-blind" problem: we have no individual labels in the source and no bag proportions in the target. The core challenge is adapting an LLP model to a new domain where the feature space and class distributions may shift significantly.

Methodology: Self-Training with Pseudo-Bags

The proposed solution is an iterative self-training algorithm designed to bridge the domain gap by creating synthetic training data in the target domain.

  1. Initial Training: A baseline LLP model (e.g., Ridge Regression or Label Regularization) is trained on the source domain bags.
  2. Salient Feature Identification: To ensure the model focuses on relevant target-domain vocabulary, Latent Dirichlet Allocation (LDA) is applied to the target domain to identify "salient" terms or topics.
  3. Pseudo-Bag Construction: The algorithm identifies features that are both salient in the target and highly predictive according to the current source-trained model. For each such feature, all target instances containing it are grouped into a "pseudo-bag." These bags are assigned a high fixed label proportion (e.g., 90% for a positive-indicative word).
  4. Iterative Refinement: The model is re-trained on these target pseudo-bags. In each iteration, the updated weights help select better features, and the number of bags is gradually increased.
  5. Ensemble Voting: To mitigate sensitivity to hyperparameters and noise in pseudo-bag labels, an ensemble of 100 models is trained with varying parameters, using majority voting for final predictions.

Experimental Setup

The approach was evaluated across five diverse text classification tasks involving significant domain shifts, such as moving from Twitter metadata to blog posts or from TV-related tweets to IMDB reviews.

Summary of datasets and tasks used for evaluation.

Results and Discussion

The proposed method, specifically the ridge-lda-ens and lr-lda-ens variants, consistently outperformed baselines across all tasks.

Classification accuracy comparison across models and tasks.

As shown in the results table, the method provided an average 11% absolute improvement in accuracy over models trained solely on the source domain. Notable gains were seen in the Politics task, where accuracy rose from 45.2% (worse than random) to 71.5%, and the Sentiment task, which improved from 52% to 65.1%.

Key Insights

  • Robustness of Bags: Adding pseudo-labeled bags proved more robust than adding individual pseudo-labeled instances. Identifying a few highly indicative features is less error-prone than labeling thousands of noisy individual instances.
  • LDA as a Bridge: LDA ensures that pseudo-bags are built around terms meaningful in the target context, preventing the model from focusing on source-specific features that may not exist in the target.
  • Correcting Sign Flips: The adaptation process allowed the model to correct "sign flips"—features that might imply one class in the source but another in the target due to context shifts.

Conclusion and Future Work

This work demonstrates that self-training with pseudo-bags is a powerful strategy for LLP domain adaptation. While effective, the method's success depends on the initial source model having some predictive power in the target domain. Future research could explore dynamic estimation of pseudo-bag proportions and the direct integration of topic modeling into the LLP objective function to further improve robustness against noise.

Summary updated

BibTeX
@inproceedings{ardehaly2016domain,
  author = {Ehsan Mohammady Ardehaly and Aron Culotta},
  title = {{Domain Adaptation for Learning from Label Proportions Using Self-Training}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure