Exploiting redundancy in large materials datasets for efficient machine learning with less data

Area of Science:

• Materials Science
• Data Science
• Machine Learning

Background:

• Large-scale materials data collection often ignores data redundancy.
• Existing datasets may contain a significant proportion of non-informative or repetitive data points.

Purpose of the Study:

• To quantify data redundancy in materials datasets.
• To investigate the impact of data redundancy on machine learning model performance.
• To explore alternative data acquisition strategies for efficient machine learning training.

Main Methods:

• Analysis of multiple large materials datasets for various properties.
• Evaluation of machine learning model performance with varying data subsets.
• Application of uncertainty-based active learning algorithms for dataset construction.

Main Results:

• Up to 95% of data can be removed from training datasets with minimal impact on in-distribution performance.
• Redundant data primarily consists of over-represented material types.
• Redundant data does not improve out-of-distribution prediction performance.
• Uncertainty-based active learning can create smaller, equally informative datasets.

Conclusions:

• The "bigger is better" approach to materials data is inefficient.
• Prioritizing data informativeness over sheer volume is crucial for effective machine learning.
• Optimized data acquisition and training strategies enhance prediction performance and robustness.