DeepCTR-Torch

Why we included this project

Ranking and response prediction over sparse categorical features usually means reimplementing the same deep architectures again. DeepCTR-Torch bundles the familiar CTR network families (factorization machines, wide & deep, DeepFM, DIN/DIEN, AutoInt, FiBiNET, xDeepFM, plus multi-task designs like ESMM and MMOE) behind one train/predict interface, so you can pit competing model ideas against each other quickly. Its feature columns hide the embedding and interaction plumbing, while the exposed core and interaction layers let you assemble a custom network when none of the shipped models fits. It also handles multi-value and sequence inputs, which covers real clickstream-style data. This is a research tool, not a production serving stack, so plan to wire up deployment yourself around the trained models.