STAR framework boosts PCVR prediction with target-aware tokenization
New KDD Cup 2026 framework blends structured tokenization and target-aware interests for sharper conversion ranking
Post-click conversion rate (PCVR) prediction remains a core ranking task for industrial recommenders, but modern models struggle to jointly handle heterogeneous non-sequential features, multi-behavior user sequences, and target-item-aware interests—all while remaining robust to high-cardinality sparse features, missing values, and train-inference inconsistencies. To address this, a team led by Yimeng Xu presents STAR (Structured Tokenization and Target-Aware Interest Representation), a practical framework built for the KDD Cup 2026 Tencent UniRec Challenge. STAR sits on a HyFormer-style multi-sequence backbone and introduces structured feature tokenization that converts raw features into discrete tokens, enabling better handling of sparse and high-cardinality signals. It also adds explicit user-item interaction tokens, target-aware sequence decoding, and a weighted user-item contrastive auxiliary objective inspired by InfoNCE, which sharpens the model's ability to capture user intent relative to a candidate item.
A key contribution is the alignment of training and inference pipelines: the authors reconstruct feature remapping tables and structural hyperparameters from the saved training configuration, reducing the train-inference skew that often plagues production recommender systems. Experiments on the UniRec challenge dataset reveal which components most reliably improve ranking AUC. The main ablation study shows a large gain from temporal context, with smaller but meaningful contributions from contrastive alignment, target-aware interest encoding, and high-cardinality sequence feature recovery. LogLoss is used as a calibration diagnostic, keeping ranking and calibration metrics distinct. For practitioners, the paper offers a modular recipe for boosting PCVR accuracy in real-world settings—particularly in large-scale ad and recommendation platforms—by improving how user behavior sequences are tokenized and how target-item relevance is injected into the ranking model.
- STAR uses structured feature tokenization to handle high-cardinality sparse features and missing values in PCVR ranking.
- Adds target-aware sequence decoding plus an InfoNCE-inspired contrastive objective for sharper user interest modeling.
- Aligns train and inference pipelines by reconstructing feature mappings from saved configs; temporal context yields the largest AUC gain.
Why It Matters
Gives recommender engineers a practical, modular framework to improve post-click conversion ranking at industrial scale.