InteractRank: Personalized Web-Scale Search Pre-Ranking with Cross Interaction Features (Pinterest, 2025)

Forum topic · 小凯 · 2026-07-05

Summary

InteractRank is a pre-ranking framework for personalized web-scale search developed by Pinterest researchers (Sujay Khandagale, Bhawna Juneja, Prabhat Agarwal, Aditya Subramanian, Jaewon Yang, Yuting Wang), presented in an April 2025 arXiv paper (arXiv:2504.06609). The work addresses the tension between ranking quality and serving latency in industrial search systems: cross-interaction features capture fine-grained query-document and user-item signals, but are typically too expensive to compute for large candidate sets during pre-ranking. InteractRank brings cross interaction signals into the early pre-ranking stage of a multi-stage ranking cascade while keeping inference costs tractable, enabling personalized retrieval at Pinterest-scale traffic. This forum entry situates the paper in the broader learning-to-rank landscape, comparing it with cascaded retrieval-rerank pipelines, generative re-ranking, and cross-encoder rerankers, and outlines typical evaluation protocols (nDCG, Recall@k, latency budgets) and engineering considerations such as index freshness, caching, and cost control. Readers should consult the original PDF for exact experimental numbers, as this entry summarizes methodology and system context rather than reproducing the paper's full results.

InteractRank: Personalized Web-Scale Search Pre-Ranking with Cross Interaction Features (Pinterest, April 2025)

Overview

Paper: InteractRank: Personalized Web-Scale Search Pre-Ranking with Cross Interaction Features (arXiv:2504.06609)
Authors / Affiliation: Sujay Khandagale, Bhawna Juneja, Prabhat Agarwal, Aditya Subramanian, Jaewon Yang, Yuting Wang — Pinterest
Category: Ranking for Search

Key Points

InteractRank targets the pre-ranking stage of a multi-stage web-scale search cascade, where the system must score millions of candidates under strict latency budgets.
It introduces cross interaction features — signals computed jointly over the query, candidate document, and user personalization context — into pre-ranking, a stage that traditionally relies on cheap, independent (non-interaction) scoring to stay fast.
The framework is designed and deployed for personalization at Pinterest scale, balancing ranking quality against inference cost.

Context in the Ranking Literature

The paper sits in the ongoing evolution of industrial learning-to-rank:

1. Classical cascade: recall (coverage) → pre-ranking (cheap filtering) → fine ranking (discriminative scoring, e.g., cross-encoders). Cross features are powerful but usually reserved for the final stage. 2. The efficiency–effectiveness trade-off: dense retrieval offers millisecond recall but is sensitive to domain shift; cross-encoders are accurate but cannot precompute document representations; generative re-ranking reduces cascade error but complicates index updates. 3. LLM-era shift: agentic pipelines now treat "how many retrieval steps and which tools" as learnable, but latency, cost, and safety remain hard production constraints.

Evaluation Considerations

Typical metrics for this problem class include nDCG@10, MRR, Recall@k, Hit@k, plus production-oriented measures such as p99 latency, per-query compute cost, and online CTR / user satisfaction. Exact numbers should be read from the original PDF; this entry summarizes the system design and positioning rather than reproducing the paper's tables.

Engineering Takeaways

| Concern | Guidance | |---------|----------| | Latency | Cascade + early stopping; cache hot queries; async reranking | | Quality | Verify offline gains translate to online CTR via interleaving experiments | | Data | Partition indexes, handle PII, version embeddings for rollback | | Cost | Distill to smaller models; mix sparse + dense scoring | | Safety | Source allowlists and output filtering for open retrieval |

Related Entries

Deep Learning to Rank in Industrial Search Engines (ACM survey)
A Generative Re-ranking Model for List-level Multi-objective Optimization (arXiv:2505.07197)
Adaptive Neural Ranking Framework: Toward Maximized Business Goal
A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE (arXiv:2403.10407)

Recommendations for Readers

1. Researchers: reproduce the core comparisons; check whether statistical significance and compute cost are reported. 2. Engineers: extract pluggable modules (encoders, rerankers) and evaluate integration cost against your existing stack. 3. Product managers: focus on user-perceivable benefits — latency, answer trustworthiness, multi-turn consistency — not just offline nDCG.