Click-through rate prediction estimates how likely a user is to click an ad or item. Identifiers like user codes get learned lookup tables, but a plain number arrives as a bare scalar, without any hint that nearby values relate. A numerical embedding turns each number into a small vector.
This work groups many candidate encoders into mechanism-based families and compares them through one shared interface, tuning each under an equal budget and testing across several datasets and downstream networks, called backbones.
DEER had the best average rank, but no encoder won everywhere: several led at least one dataset and backbone pairing. Feeding inputs at their original scale instead of standardized scores flipped the winner in 11 of 27 settings.