Polynomial feature interaction transforms input features into powers and products, allowing a linear model to represent nonlinear relationships in the original feature space. For features a and b, degree-two expansion produces 1, a, b, a², ab, and b²; products such as ab are interaction terms because they represent combined effects. section 3.4.2
The workflow transforms training data, fits a linear model in the expanded space, and applies the same transformation before predicting new data. Higher degrees increase flexibility but can cause feature-count growth, computation, and overfitting. section 3.4.2
Use polynomial interactions with attention to representation and evaluation:
- Tune the transformation or its hyperparameters through cross-validation within the training set, not by repeatedly consulting the test set. section 3.4.2
- Treat molecular fingerprints cautiously: multiplying binary bits often adds redundant information, while sparse interactions can be nearly always zero. section 3.4.4
hERG Blocking Prediction
For hERG blocking prediction, polynomial features derived from molecular fingerprint bits were not informative in randomized search. A final regularized linear model achieved a test F1 score of 0.866, compared with cross-validated F1 of 0.881. section 3.4.4


