Improving molecular machine learning through adaptive subsampling with active learning
Data subsampling is an established machine learning pre-processing technique to reduce bias in datasets. However, subsampling can lead to the removal of crucial information from the data and thereby decrease performance. Multiple different subsampling strategies have been proposed, and benchmarking is necessary to identify the best strategy for a specific machine learning task. Instead, we propose to use active machine learning as an autonomous and adaptive data subsampling strategy. We show that active learning-based subsampling leads to better performance of a random forest model trained on Morgan circular fingerprints on all four established binary classification tasks when compared to both training models on the complete training data and 16 state-of-the-art subsampling strategies. Active subsampling can achieve an increase in performance of up to 139% compared to training on the full dataset. We also find that active learning is robust to errors in the data, highlighting the utility of this approach for low-quality datasets. Taken together, we here describe a new, adaptive machine learning pre-processing approach and provide novel insights into the behavior and robustness of active machine learning for molecular sciences.