Deep Position-wise Interaction Network for CTR Prediction

The Meituan advertising team extends its previous work on bias mitigation by focusing on position bias in click‑through‑rate (CTR) prediction. Building on the framework presented in their KDD Cup 2020 paper, they propose a novel Deep Position-wise Interaction Network (DPIN) that jointly models ads and positions.

Model Architecture

DPIN consists of three modules:

Base Module : an embedding‑MLP pipeline that encodes user, context, and candidate‑ad features.

Deep Position-wise Interaction Module : extracts position‑specific behavior sequences, aggregates them with attention, applies a non‑linear fully‑connected layer, and further refines the representations with a stack of Transformer blocks.

Position-wise Combination Module : combines the ad representation from the Base Module with the position representation from the interaction module via a non‑linear layer to produce a position‑aware CTR estimate.

The overall loss is binary cross‑entropy, and training uses real position labels.

Experiments

Offline evaluations on billions of Meituan search‑keyword ad impressions (both regular traffic and random‑exploration traffic) show that DPIN improves AUC by up to 2.98% and introduces a new metric, PAUC (Position‑wise AUC), which better reflects bias‑corrected ranking quality. Service‑side latency grows only modestly with candidate‑ad count because the heavy sequence modeling is confined to the position‑wise module.

Online A/B tests confirm the gains: DPIN raises CTR by 2.25% and RPM (revenue per mille) by 2.15% compared with the production baseline.

Conclusion

DPIN effectively mitigates position bias while maintaining high service performance, and its design—deep non‑linear interaction among position, context, and user—can be extended to other combinatorial ranking problems in advertising.