AdsRank
ranking
pytorch
c++
machine-learning
A three-stage CTR ranking system on Criteo: two-tower + FAISS retrieval, DeepFM ranking, and ONNX Runtime serving in C++.
Overview
AdsRank is a small version of the retrieve-then-rank funnel used in production ad systems, trained on a 5-million-impression sample of the Criteo click logs.
- Retrieval. Two-tower embeddings with FAISS narrow the full catalog to a candidate set.
- Ranking. A PyTorch DeepFM model scores those candidates for click-through rate.
- Serving. The ranker is exported to ONNX and served by a C++ inference benchmark.
Results
| Stage | Result |
|---|---|
| Ranking | Held-out CTR AUC 0.782 → 0.801, log loss −4.2% |
| Retrieval + ranking | 96% of an exhaustive ranker’s top-10 kept, scoring 90% fewer items |
| Serving | p95 CPU latency 42 ms → 24 ms at a fixed batch size |
Ranking model
DeepFM learns embeddings for the categorical features and models interactions between them. Ablations over those interactions showed which feature crosses mattered.
Serving in C++
After exporting to ONNX, latency improvements came from tuning ONNX Runtime thread counts and reusing input buffers so each request doesn’t allocate new ones.
Stack
Python · C++ · PyTorch · FAISS · ONNX Runtime · Linux