Random feature selection, also called feature subsampling, chooses only a random subset of input features when constructing each decision-tree split. In a Random Forest classifier, different trees therefore see different feature choices as well as different training-data samples. This decorrelates the trees, increases ensemble diversity, and helps reduce variance and overfitting while retaining the low bias of decision trees. section 5.3.2 section 5.5
A practical Random Forest workflow uses feature subsampling alongside ensemble evaluation:
- Represent molecules as input features such as fingerprint bits. section 5.3.2
- Train many decision trees, allowing each split to consider only a random feature subset. section 5.3.2
- Combine the trees’ predictions to obtain a more robust classifier. section 5.3.2
- Use out-of-bag samples to estimate performance without a separate validation set, especially when data are limited. section 5.3.2
CYP3A4 inhibition prediction
A Random Forest trained on molecular fingerprints achieved a validation PR AUC of 0.837, while its out-of-bag PR AUC was 0.806. Feature subsampling contributes to the ensemble’s diversity; feature importance can then rank influential fingerprint bits.

