Epsilon-insensitive loss ignores small prediction errors inside an epsilon margin and penalizes larger errors linearly. This makes regression less sensitive to noisy experimental solubility labels. section 4.3.3
Use it when:
- A descriptor-based regression problem contains noisy measurements or outliers. section 4.3.3
- A linear model operates in an expanded feature space, such as polynomial molecular descriptors. section 4.3.3
- Small errors are practically unimportant and should not dominate optimization. section 4.3.3
Solubility regression pipeline
An SGDRegressor uses loss="epsilon_insensitive" with L2 regularization, early stopping, inverse-scaling learning rate, imputation, polynomial features, and standardization. The resulting model achieved a test RMSE of 1.2783, below ESOL's 1.3787 baseline. section 4.3.3
