Support vector regression (SVR) is a supervised regression method that finds a hyperplane while allowing a margin of error around predicted values. Its epsilon-insensitive loss ignores small errors, making it useful for noisy experimental solubility labels. section 4.3.3 section 4.5
For descriptor-based molecular data, SVR remains linear in its feature space but can represent nonlinear relationships after feature expansion. This makes it a practical option for small-to-medium datasets with engineered molecular descriptors. section 4.3.3
A practical SVR workflow includes:
- Create molecular descriptors, optionally expand them with PolynomialFeatures, and scale the resulting features. section 4.4.3
- Train with a regularized estimator such as SGDRegressor using epsilon-insensitive loss, a learning-rate schedule, and optional early stopping. section 4.3.3
- Tune regularization, polynomial degree, and epsilon-tube width, then inspect validation curves for underfitting and overfitting. section 4.4.3
- Compare test-set RMSE with meaningful baselines rather than interpreting the score alone. section 4.3.3
Molecular solubility prediction
An SVR model trained on engineered molecular descriptors achieved a validation RMSE of 1.3791. The selected model reached a test RMSE of 1.2783, outperforming the ESOL baseline at 1.3787 and the y-scrambled baseline at 2.4585. Its pipeline combined median imputation, degree-two polynomial features, standardization, L2 regularization, epsilon-insensitive loss, inverse-scaled learning, and early stopping. section 4.3.3

