Research Paper#Materials Science, Machine Learning, Multi-Task Learning🔬 ResearchAnalyzed: Jan 3, 2026 19:40
MTL Failure in Alloy Property Prediction: Data Imbalance and Task Independence
Published:Dec 28, 2025 01:52
•1 min read
•ArXiv
Analysis
This paper investigates the conditions under which Multi-Task Learning (MTL) fails in predicting material properties. It highlights the importance of data balance and task relationships. The study's findings suggest that MTL can be detrimental for regression tasks when data is imbalanced and tasks are largely independent, while it can still benefit classification tasks. This provides valuable insights for researchers applying MTL in materials science and other domains.
Key Takeaways
- •MTL can negatively impact regression tasks when data is imbalanced and tasks are independent.
- •MTL can improve classification performance, especially recall, even with data imbalance.
- •Careful consideration of data characteristics and task relationships is crucial when applying MTL.
Reference
“MTL significantly degrades regression performance (resistivity $R^2$: 0.897 $ o$ 0.844; hardness $R^2$: 0.832 $ o$ 0.694, $p < 0.01$) but improves classification (amorphous F1: 0.703 $ o$ 0.744, $p < 0.05$; recall +17%).”