Easy-to-use,Modular and Extendible package of deep-learning based CTR models .
8k
Stars
2.2k
Forks
62
Open issues
14
Contributors
AI Analysis
DeepCTR is a Python library for building and deploying deep learning models for click-through rate (CTR) prediction, a specialized task in recommendation systems and online advertising. It provides modular, reusable components (attention mechanisms, embedding layers, feature interactions) and pre-built models (DeepFM, xDeepFM, AutoInt, DIEN) compatible with TensorFlow 1.15 and 2.x. Best suited for practitioners in e-commerce, ads, and recommendation platforms who need to rapidly prototype and...
Inferred from signals mentioned in the README (tests, CI, type safety) — not a review of the actual code.
AI's overall editorial judgment — not an average of the bars above, can weigh other factors too.
DeepCTR: A modular TensorFlow library for click-through rate prediction models used in ad and recommendation systems
DeepCTR is a Python library built on TensorFlow/Keras that implements a wide range of peer-reviewed deep learning CTR models (DeepFM, Wide&Deep, DCN, AFM, NFM, and many others) in a unified, composable interface. It targets ML practitioners and researchers working on ad click prediction, recommender systems, and related ranking problems. With 8K+ stars, 2.2K forks, PyPI downloads in the millions, and an active Chinese ML community around it, it has meaningful real-world traction. It lowers the barrier to experimenting with and deploying research-grade CTR architectures.
Started in October 2017 alongside the wave of deep learning adoption in industrial recommendation systems. It predates many equivalent libraries and has grown incrementally as new CTR papers emerged, adding both TF1/TF2 compatibility and a companion PyTorch variant (DeepCTR-Torch).
Growth was driven primarily by the Chinese ML and recommender systems community, fueled by Zhihu articles, academic course references, and the relative scarcity of clean reference implementations for CTR models. Star accumulation has plateaued at a steady but slow rate (3 stars/week observed), suggesting a mature, stable project rather than a fast-growing one. The ecosystem expanded with DeepMatch and DeepCTR-Torch as companion projects.
PyPI download badge is present (pepy.tech link), suggesting substantial cumulative downloads. The Chinese ML community engagement (Zhihu introduction, WeChat discussion group) indicates real practitioner use. Adoption not independently verified at large-scale production deployments, but download volume and fork count (2,223) strongly imply practical usage beyond casual experimentation.
Appears to follow a Keras-native layered design where individual model components (embedding layers, interaction layers, DNN layers) are implemented as reusable Keras layers, then composed into full models. Likely uses a shared base structure so all models expose the same fit/predict API. Estimator support for distributed training is explicitly mentioned, suggesting a dual-interface design.
CI badges for both TF1 and TF2 pipelines are present, and a Codecov badge is shown — coverage appears to be actively tracked. Exact coverage percentage is not stated in the README excerpt, but the presence of Codecov integration and dual CI workflows suggests meaningful automated testing.
Last push was 2026-06-24, three days before the evaluation date — the repository is actively maintained. CI workflows are running. Issue tracker is referenced. Maintenance appears healthy for a mature library, though feature velocity is likely low given the project's age and stability.
ADOPT IF: you need clean, well-tested reference implementations of classic and modern CTR models (DeepFM, DCN, etc.) in TensorFlow/Keras and want a stable, documented library for experimentation or baseline comparisons. AVOID IF: you are starting a new project on PyTorch (use DeepCTR-Torch instead), or need cutting-edge 2024-2026 model architectures — the model list appears anchored to the 2015–2022 paper era. MONITOR IF: you depend on TensorFlow long-term, as TF's own trajectory affects this library's relevance; also monitor if PyTorch-based CTR libraries absorb the community.
Independent dimensions
Mainstream potential
3/10
Technical importance
7/10
Adoption evidence
6/10
- TensorFlow's declining mindshare relative to PyTorch could reduce the relevance of the TF-based DeepCTR over time, shifting practitioners to DeepCTR-Torch.
- Model list may not keep pace with the latest CTR research (e.g., transformer-based and LLM-augmented ranking models from 2023 onward), risking staleness for cutting-edge use cases.
- Primarily maintained by a single core author, which creates a bus-factor risk for long-term sustainability despite current activity.
- TensorFlow version compatibility (TF1.15 through TF2.x) introduces ongoing maintenance burden as TF deprecates internal APIs, as noted in the README's warnings about private keras APIs.
- Community is heavily concentrated in the Chinese ML ecosystem; English-language documentation and international community engagement appear secondary, which may limit broader adoption.
DeepCTR will likely remain a stable reference library for CTR practitioners on TensorFlow, with slow but continued maintenance. It is unlikely to regain rapid growth but will persist as a valued niche resource, especially in Chinese industry and academia.
Newsletter
Get analyses like this every Monday
Free weekly digest of the most interesting open-source discoveries.
Languages
Information
- Language
- Python
- License
- Apache-2.0
- Last updated
- 1w ago
- Created
- 107mo ago
- Analyzed with
- anthropic/claude-haiku-4-5
Stars over time
Contributors over time
Top 100 contributors only — repos with more will plateau at 100.
Open issues
请问能否帮忙解决一下，使用DEEPFM会存在内存泄漏的问题
The following Variables were used a Lambda layer's call，BatchNormalization
mmoe训练模型，测试集ctr和cvr的auc完全相等。
Value error when loading saved DIEN model
EDCN应该每次传入cross_in和初始第0层的cross_in，否则不符合DCN本身数学推导
Top contributors
Recent releases
Similar repos
nl8590687/ASRT_SpeechRecognition
ASRT is a deep learning-based Chinese speech recognition system built with...
datawhalechina/torch-rechub
Torch-RecHub is a PyTorch-based framework for building recommendation systems,...
datawhalechina/fun-rec
A comprehensive educational resource teaching recommendation system algorithms,...
| Repository | Stars | Week Δ | Language | Score | Updated |
|---|---|---|---|---|---|
|
shenweichen/DeepCTR
this
|
8k | +3 | Python | 7/10 | 1w ago |
|
nl8590687/ASRT_SpeechRecognition
|
8.4k | — | Python | 7/10 | 3mo ago |
|
datawhalechina/torch-rechub
|
1.2k | — | Jupyter Notebook | 7/10 | 3d ago |
|
datawhalechina/fun-rec
|
7.2k | — | Python | 7/10 | 2w ago |
 wang-xinyu/tensorrtx
|
7.8k | — | C++ | 8/10 | 2w ago |
The author's own PyTorch port. Many practitioners now prefer PyTorch for research, making DeepCTR-Torch a direct sibling that may see more new adoption while DeepCTR (TF) serves existing TF-based pipelines.
Microsoft's library covers a broader scope including collaborative filtering, NLP-based rec, and CTR models, with stronger enterprise documentation. DeepCTR is more focused and arguably simpler to use for pure CTR model experimentation.
Google's official TF library for retrieval and ranking. More idiomatic TF2 integration and better long-term TF compatibility, but covers fewer specific CTR paper implementations than DeepCTR.
A newer, more research-oriented CTR benchmarking framework with broader model coverage and reproducibility focus. May appeal more to researchers; DeepCTR remains more accessible for practitioners needing quick deployment.
PyTorch-based library with similar goals for Chinese-speaking practitioners. Growing in the same community space but younger and with less accumulated documentation.
