DeepCTR

Why we included this project

For teams building ad bidding, feed ranking, or any click-prediction pipeline, this library compresses a decade of published CTR research into one importable package. It implements the well-known families — Wide & Deep, DeepFM, xDeepFM, Deep Interest Network, DIN/DIEN, and many others — behind a single tf.keras-style fit/predict interface, so you can swap model architectures without rewriting your data plumbing. Because every model is built from reusable feature-interaction and attention layers, you can also assemble custom architectures when a paper's approach doesn't map cleanly onto an off-the-shelf option. It runs on both TensorFlow 1.15 and 2.x and offers an estimator path for large-scale distributed training, which makes it practical for moving from a quick offline experiment to a deployed ranking model.