Transfer Learning with Active Queries from Source Domain

Sheng-Jun Huang, Songcan Chen

Abstract

In many real-world transfer learning scenarios, labeled data is scarce in both source and target domains, and expert annotation is only available for the source domain. This paper introduces TLAS (Transfer Learning with Active queries from Source), a framework that jointly optimizes the selection of informative source instances for labeling and the weighting of these instances to bridge the domain gap. Experimental results across sentiment analysis and text categorization tasks demonstrate that TLAS significantly outperforms standard active learning and domain adaptation baselines.

What Problem Does This Paper Address?

Traditional transfer learning assumes a labeled source domain helps an unlabeled target domain. Conversely, active learning assumes an oracle can label the most informative data within the domain of interest. However, "double scarcity"—where labels are missing in both domains—and "oracle asymmetry"—where experts are only available for the source domain—present a unique challenge.

The TLAS framework: Active queries are sent to an oracle in the source domain to improve performance in the target domain.

The core challenge is: How can we strategically query labels in a source domain to maximize model performance on a different target domain? This is critical when target data is private or when experts for the target domain are unavailable.

Methods

The TLAS framework treats domain adaptation and active selection as a joint optimization problem. It integrates three primary components:

  1. Distribution Matching via MMD: Maximum Mean Discrepancy (MMD) is employed to minimize the distance between the weighted source domain and the target domain, while also ensuring selected samples are representative of the broader unlabeled source data.
  2. Importance Weighting ($\beta$): To handle covariate shift, the model learns weights ($\beta$) for source instances, effectively re-weighting the source distribution to match the target.
  3. Active Selection ($\alpha$): A selection vector ($\alpha$) identifies which source instances to query. This selection is driven by both representativeness (via MMD) and an Uncertainty Term, prioritizing instances where the model's current confidence is low.

The resulting biconvex optimization problem is solved using Alternating Quadratic Programming, iteratively updating weights and selection criteria until convergence.

Experimental Setup

The researchers evaluated TLAS on 15 datasets:

  • Sentiment Analysis: 12 domain pairs from Amazon reviews (Books, DVD, Electronics, Kitchen).
  • Text Categorization: 3 classification tasks from the Reuters-21578 collection (Orgs, People, Places).

Experiments simulated label scarcity by providing only 10% initial labels for both domains.

Key Results

TLAS demonstrated consistent superiority across all evaluated tasks.

Accuracy comparison across 12 sentiment analysis domain pairs as the number of queried instances increases.
  • Efficiency: In sentiment analysis, TLAS often achieved higher accuracy with 100-200 queries than baselines (Random, Uncertainty, MPAL) achieved with 500.
  • Joint Optimization Benefit: A version of the model with fixed weights (TLAS-b) performed well, but the full TLAS framework—which updates weights iteratively—was significantly better, highlighting the synergy between selection and adaptation.
  • Robustness: The performance gap remained significant even as the initial labeled data ratio in the target domain was increased from 10% to 50%.

Limitations and Future Work

While highly effective, the authors noted instances of "negative transfer" in specific text categorization tasks where excessive source data eventually degraded performance. Future research directions include extending the framework to Multi-Source Transfer Learning and exploring alternative active learning strategies like "query-by-committee."

Summary updated

Abstract

To learn with limited labeled data, active learning tries to query more labels from an oracle, while transfer learning tries to utilize the labeled data from a related source domain. However, in many real cases, there is very few labeled data in both source and target domains, and the oracle is unavailable in the target domain. To solve this practical yet rarely studied problem, in this paper, we jointly perform transfer learning and active learning by querying the most valuable information from the source domain. The computation of importance weights for domain adaptation and the instance selection for active queries are integrated into one unified framework based on distribution matching, which is further solved with alternating optimization. The effectiveness of the proposed method is validated by experiments on 15 datasets for sentiment analysis and text categorization.
BibTeX
@inproceedings{huang2016transfer,
  author = {Sheng-Jun Huang and Songcan Chen},
  title = {{Transfer Learning with Active Queries from Source Domain}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure