Joint Optimization of Ranking and Calibration (JRC) for CTR Prediction
The Joint Optimization of Ranking and Calibration (JRC) model introduces a two‑logit generative‑discriminative architecture that jointly minimizes LogLoss for calibration and a listwise ranking loss, delivering superior GAUC and CTR performance across Alibaba’s display‑ad system, especially for sparse long‑tail users, while remaining simple to train and deploy.
We propose a hybrid generative/discriminative model called Joint Optimization of Ranking and Calibration (JRC) that simultaneously improves ranking ability and calibration accuracy for click‑through‑rate (CTR) prediction in advertising.
Traditional industrial models optimize pointwise LogLoss for calibration, while ranking metrics such as AUC/GAUC are optimized separately. JRC increases model degrees of freedom by outputting two logits (click and non‑click) and designs a logits interaction that replaces the single‑logit sigmoid with a softmax, preserving calibration via LogLoss and enhancing ranking via a listwise loss that pushes positive logits above negatives within the same session.
The final loss combines the calibrated LogLoss term and the listwise ranking term, weighted by a hyper‑parameter that is empirically insensitive. The approach can be interpreted as optimizing a mixed generative/discriminative objective, where the discriminative part enforces calibration and the generative part expands the logits to improve ranking.
Extensive offline experiments on Alibaba’s display‑ad dataset show that JRC outperforms pointwise, combined (pointwise + pairwise/listwise), and multi‑task baselines: it achieves higher GAUC while maintaining or improving LogLoss. Online A/B tests during the 618 shopping festival demonstrate significant gains (RPM +2.3 %, CTR +4.4 %, PPC ‑2.1 %, ROI +2.8 %) and a GAUC increase of 0.3 %.
Analysis reveals that JRC especially benefits sparse‑long‑tail users, where the additional generative supervision provides stronger signals. The model has been fully deployed across all Alibaba‑Mama display‑ad scenarios.
Future work includes extending the two‑logit formulation to deeper generative models, semi‑supervised learning, and uncertainty estimation.
Alimama Tech
Official Alimama tech channel, showcasing all of Alimama's technical innovations.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.