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 web-scale personalized search developed at Pinterest, published as arXiv:2504.06609 in April 2025. The system addresses a classic industrial search problem: cross-interaction models (which jointly encode query and candidate features) deliver strong ranking quality but are too expensive to run over millions of candidates. InteractRank brings cross interaction features into the pre-ranking stage, enabling lightweight early-stage scoring that better preserves the candidates that a downstream ranker would prefer. The work targets personalized search at web scale, where retrieval and pre-ranking must filter a huge candidate pool within tight latency budgets. This forum post indexes the paper alongside related industrial ranking research, including deep learning-to-rank in search engines, generative re-ranking for list-level multi-objective optimization, cross-encoder versus LLM reranking comparisons, and adaptive neural ranking frameworks. The post includes a meta-information table, positioning within ranking-for-search literature, and engineering notes on latency, cost, and evaluation constraints. Full quantitative results should be verified against the original PDF at arxiv.org/abs/2504.06609.

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

Overview

This entry indexes a paper from Pinterest on pre-ranking for personalized web-scale search:

InteractRank: Personalized Web-Scale Search Pre-Ranking with Cross Interaction Features

Authors / Affiliation: Sujay Khandagale, Bhawna Juneja, Prabhat Agarwal, Aditya Subramanian, Jaewon Yang, Yuting Wang (Pinterest)
Source: https://arxiv.org/abs/2504.06609
Type: Academic paper
Section: Ranking for Search

Problem and Motivation

Large-scale search, recommendation, and personalization systems have long faced tension between ranking quality and serving efficiency. Traditional cascaded pipelines separate retrieval, pre-ranking, and final ranking, which limits how rich the interaction between query and candidate features can be in early stages. Cross interaction features — where query and candidate signals are jointly modeled — improve relevance but are typically too costly to compute for every candidate at web scale.

InteractRank addresses this by introducing cross interaction features into the pre-ranking stage, aiming to make early-stage filtering more consistent with the preferences of the downstream ranker while meeting strict latency budgets in personalized search.

Context in the Ranking Literature

Neural ranking has evolved from BM25 and dual-tower dense retrieval to cross-encoders, late interaction, and generative/LLM-based approaches, each balancing effectiveness, efficiency, and maintainability. Cross-encoders achieve high accuracy but cannot precompute document representations; dense retrieval enables millisecond-level recall but is sensitive to domain shift and long-tail queries. Pre-ranking methods like InteractRank sit between these extremes: they approximate richer cross-interaction scoring at a fraction of the cost.

Related Entries

Deep Learning to Rank in Industrial Search Engines, Recommender Systems (10.1145/3797895)
Multi-Objective Recommendation in the Era of Generative AI: A Survey (arXiv:2506.16893)
A Generative Re-ranking Model for List-level Multi-objective Optimization (arXiv:2505.07197)
A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE (arXiv:2403.10407)
Accelerating Listwise Reranking: Reproducing and Enhancing FIRST, SIGIR (10.1145/3726302.3730287)
Adaptive Neural Ranking Framework: Toward Maximized Business Goals (10.1145/3589334.3645605)

Engineering Checklist

| Item | Question | Suggestion | |------|----------|------------| | Latency | What is the p99 budget? How many retrieval steps? | Cascading with early stopping, caching popular queries, async reranking | | Quality | Do offline gains translate to online CTR/satisfaction? | Interleaving experiments, manual audits | | Cost | Per-query compute and GPU usage? | Small-model routing, distillation, hybrid sparse+dense retrieval | | Data | PII in training/indexing data? | Partitioned indices, anonymization, versioned embeddings |

Notes for Readers

Researchers: Reproduce core comparisons and check for statistical significance and compute cost reporting.
Engineers: Evaluate how pluggable components (pre-ranker, ranker) integrate with existing serving stacks.
Product managers: Focus on user-perceivable benefits (latency, answer quality) rather than only offline nDCG.

Specific quantitative results should be verified against the original PDF at arxiv.org/abs/2504.06609.