reczoo/FuxiCTR

FuxiCTR – Open‑source Click‑Through‑Rate (CTR) Prediction Library

What it is

• A Python package (available on PyPI) that bundles a large collection of state‑of‑the‑art CTR prediction models. It targets industrial‑scale tasks such as online advertising, recommender systems, and sponsored search.
• Works with both PyTorch 2.5+ and TensorFlow 2.21+, so you can pick the deep‑learning framework you prefer.

Why it matters

• CTR prediction is a core component of many revenue‑generating services. Researchers need a reproducible benchmark, and engineers need a plug‑and‑play implementation that can be tuned and extended quickly. FuxiCTR provides both.

Key design points

Feature	What it gives you
Configurable	Data preprocessing pipelines and model architectures are modular; you can assemble them via simple YAML/JSON configs instead of writing code.
Tunable	Hyper‑parameter search is baked in – you can launch automatic tuning runs by editing the same config file.
Reproducible	All published benchmarks (the “Model Zoo”) can be rerun with a single command, ensuring results match the papers.
Extensible	Adding a new model only requires implementing a PyTorch or TensorFlow module; the rest of the pipeline (data handling, training loop, evaluation) is reused.

Model zoo

• The repository ships 41+ ready‑to‑use models ranging from classic Linear Regression (LR) and Factorization Machines (FM) to recent graph‑based and attention‑based architectures such as FiGNN, AutoInt, and WuKong.
• Each entry lists the original conference/journal, a link to the paper, and the framework(s) it supports (torch, tf, or both).
• Models are organized into two logical groups:
  1. Feature Interaction Models – pure CTR predictors that focus on learning high‑order interactions among sparse features.
  2. Behavior Sequence Modeling – models that incorporate user click/interaction histories (e.g., DIN, DIEN, BST).

Getting started

# Install the latest release
pip install fuxictr

# Verify the environment (requires Python ≥3.10)
python -c "import fuxictr, torch, tensorflow as tf; print(fuxictr.__version__)"

• After installation, clone the repo to get example configs and data scripts:

git clone https://github.com/reczoo/FuxiCTR.git
cd FuxiCTR

• Run a benchmark (e.g., the DeepFM model) with a single command:

python run_experiment.py --config configs/DeepFM.yaml

The config file lets you specify the dataset, feature processing, model hyper‑parameters, and the number of tuning trials.

Who should use it

• Researchers who want to compare a new CTR idea against a comprehensive set of baselines without re‑implementing each paper.
• Practitioners in ad tech or recommendation platforms who need a production‑ready, configurable CTR engine that can be swapped between PyTorch and TensorFlow.
• Students looking for a concrete codebase to study how modern recommendation models are built and evaluated.

License & community

• Distributed under the Apache‑2.0 license (per the badge in the README).
• The project shows active interest (badge for downloads, star‑counter image) and provides links to the original papers for each model, making it easy to trace back to the research.

---

All details above are taken directly from the repository’s README; no external assumptions have been added.