Problem
An ad platform has to predict how an ad performs for advertisers with different goals: views, clicks, visits or sales. The older setup relied on heavy feature engineering and trained a separate model for each event, so it could not optimize several goals together. Late events like sales are rare, the training set has billions of rows, and bidding allows up to a billion inferences per second with serving latency under 10 ms.
Example from the article: Criteo introduced multi-task deep models that cover view, click, visit and sale in one model, and reports a 3% uplift in Return on Ad Spend (ROAS) for participating advertisers.
Solution in brief
One deep model shares an embedding layer and a residual MLP backbone, and has one output head per funnel event. Rare events learn from representations trained on frequent ones. The model is fine-tuned continuously on fresh data, and a post-training calibration step fixes the offline/online shift during A/B tests.
Questions this case answers
- How to predict several nested user actions with one model when later actions are much rarer than earlier ones
- How to keep a large model fresh on a continuously changing set of categories without full retraining
- How to roll out a model whose predictions are calibrated on data produced by the previous model, when the new model changes the traffic it later sees
- How to predict several funnel events (view, click, visit, sale) for an ad with one production model instead of separate models?
Tools and methods
- Multi-task deep model — embedding layer, MLP backbone, one head per event.
- Continual fine-tuning — daily updates on new data, with periodic retraining from scratch.
- Post-training calibration — corrects predictions on traffic acquired by the tested model.
- PyTorch, Ray Data, Ray Train — training and distributed data loading.
- ONNX, ONNX Script, Triton, TensorRT — export, preprocessing and GPU serving.
Google or a one-time email code. No password.