Improving Protein-Protein Interaction Article Identification
Author Information
Author(s): Tsai Richard Tzong-Han, Hung Hsi-Chuan, Dai Hong-Jie, Lin Yi-Wen, Hsu Wen-Lian
Primary Institution: Yuan Ze University and Academia Sinica
Hypothesis
Can likely-positive and unlabeled data improve the identification of protein-protein interaction articles?
Conclusion
Integrating unlabeled and likely labeled data enhances the performance of a protein-protein interaction text classification system.
Supporting Evidence
- Adding likely-labeled data generally increases AUC by 3~6%.
- Our final model achieves an F-measure and AUC 2.9% and 5.0% higher than the top-ranking system.
- The hierarchical model has higher F-measures while the mixed model has higher AUCs.
Takeaway
This study shows that using both positive and negative data helps computers better find articles about how proteins interact with each other.
Methodology
The study used a machine-learning-based classifier trained on labeled and unlabeled data to identify relevant articles.
Potential Biases
The model may be biased towards positive predictions if the training data is imbalanced.
Limitations
The reliance on the quality of the labeled data and the potential noise in the unlabeled data may affect the model's performance.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website