Abstract
Extracting protein-protein interactions (PPI) from biomedical literature is a critical task for biological knowledge discovery, yet it is hampered by the high cost of expert data labeling. This paper presents ActTrAdaBoost, a novel framework that integrates transfer learning with active learning. By employing a "relative distribution" weighting method (RDTrAdaBoost), the framework effectively mitigates negative transfer from source domains. Experimental results on gold-standard corpora demonstrate that ActTrAdaBoost significantly outperforms traditional supervised and transfer learning methods, particularly when labeled target data is extremely scarce.
Introduction
The rapid growth of biomedical literature necessitates automated methods for Protein-Protein Interaction Extraction (PPIE). While supervised machine learning models have shown promise, their performance depends on large volumes of high-quality labeled data. Transfer learning offers a potential solution by leveraging existing labeled datasets from related domains. However, differences in data distributions between source and target domains often lead to "negative transfer," where the source data degrades model performance. The authors aim to bridge this gap by making transfer learning more robust and data-efficient.
The ActTrAdaBoost Framework
To address these challenges, the authors propose the ActTrAdaBoost framework, which combines an improved transfer learning algorithm with an active learning loop.
The framework consists of two primary components:
- RDTrAdaBoost (Relative Distribution TrAdaBoost): Standard TrAdaBoost treats all source instances equally at initialization. RDTrAdaBoost improves upon this by initializing weights based on the relative distribution ($P_T(x)/P_S(x)$). This approach prioritizes source instances that closely resemble the target domain, filtering out noise and reducing the risk of negative transfer.
- Active Learning Integration: The framework incorporates an active learning loop using a Support Vector Machine (SVM) to identify samples with high uncertainty. When the model's confidence is low or the training set is insufficient, it queries a human expert for labels. For high-confidence predictions, the model uses its own output to expand the training set autonomously.
Experimental Evaluation
The framework was evaluated using three benchmark biomedical corpora: AIMed, IEPA, and HPRD50. The authors utilized Kullback–Leibler (KL) divergence to confirm significant distributional differences between these datasets, justifying the need for a robust transfer learning approach.
Results and Discussion
ActTrAdaBoost consistently outperformed baseline models, including standard SVMs and the original TrAdaBoost algorithm. The framework's strengths are most apparent in data-sparse scenarios.
Key findings include:
- Efficiency at the Edge: With only 2% of the target data labeled by experts, ActTrAdaBoost achieved a 26.69% improvement in F-score over standard TrAdaBoost on the IEPA dataset.
- Robustness to Negative Transfer: RDTrAdaBoost successfully maintained performance even when source and target domains were highly dissimilar (e.g., using HPRD50 as a source for IEPA), where traditional methods typically falter.
Conclusion
ActTrAdaBoost provides a powerful solution for PPIE in scenarios where expert annotation is limited. By combining distribution-aware transfer learning with active learning, the framework maximizes the utility of both existing source data and limited target labels. Future work may involve integrating deep learning architectures to automate feature extraction and further enhance performance.
Summary updated