Reproduce “STAR: Structured Tokenization and Target-Aware Interest Representation for PCVR Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yimeng Xu, Ruihao Zhang, Yingqi Song, Ying Jiang, Lan Ma
STAR predicts post-click conversion for candidate items by tokenizing user and item features, preserving rare signals, distinguishing real zeros from padding, and conditioning behavior histories on the candidate. It adds target-aware decoding and weighted InfoNCE to a HyFormer-style backbone. On the KDD Cup 2026 Tencent UniRec Challenge test set, the full system reached AUC 0.836546 versus 0.819961 baseline, offline only.
Key insights
- Temporal context features, robust missing-value handling, and a contrastive learning objective are crucial for reliably improving PCVR prediction accuracy (AUC).
- A DIN-style query decoder is effective in generating target-aware user interests, demonstrating an advantage over a general cross-attention mechanism for this task.
- Meticulous train-inference consistency and efficient utilization of high-cardinality sparse features are vital for achieving robust performance in real-world industrial recommender systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.