A Universal Framework for Compressing Embeddings in CTR Prediction (arXiv 2502.15355)

Forum topic · 小凯 · 2026-07-05

Summary

This forum post catalogs a February 2025 arXiv paper, 'A Universal Framework for Compressing Embeddings in CTR Prediction' (arXiv:2502.15355), authored by Kefan Wang, Hao Wang, Kenan Song, Wei Guo, Kai Cheng, Zhi Li, and colleagues. The paper addresses a well-known pain point in large-scale search, recommendation, and advertising systems: embedding tables in CTR (click-through rate) prediction models consume enormous memory as the item catalog and feature vocabulary grow. The post frames the work within the broader evolution of embedding models—from matrix factorization and deep CTR models to Transformer-based sequence models—and situates it in a curated reading list alongside related resources such as BGE-M3, Arctic-Embed 2.0, and BitNet text embeddings. Note that the post itself is largely a template-based summary: most sections (method details, experimental numbers, conclusions) are placeholders that defer to the original PDF rather than quoting concrete results. Readers interested in the actual compression technique, benchmarks, and accuracy-memory trade-offs should consult the arXiv abstract and paper directly.

A Universal Framework for Compressing Embeddings in CTR Prediction (arXiv 2502.15355, Feb 2025)

Overview

This post indexes a February 2025 arXiv paper on a universal framework for compressing embeddings in CTR (click-through rate) prediction.

| Field | Content | |---|---| | Title | A Universal Framework for Compressing Embeddings in CTR Prediction | | Authors | Kefan Wang, Hao Wang, Kenan Song, Wei Guo, Kai Cheng, Zhi Li, et al. (9 authors total) | | Published | February 2025 (arXiv) | | Link | https://arxiv.org/abs/2502.15355 | | Type | Academic paper | | Section | Embedding models |

Background

In large-scale search, recommendation, and personalization systems, embeddings have long faced challenges around efficiency, scalability, and user-intent understanding. Embedding tables in CTR prediction models are a major memory bottleneck: as the item catalog and feature vocabulary grow, storing full-precision per-feature embeddings becomes impractical. This paper is positioned within that problem space, proposing a general framework for embedding compression in CTR models.

> Editor's note: The original forum post is largely template-generated. Sections on method architecture, experimental results, and conclusions are placeholders that do not quote concrete figures from the paper. Specific quantitative results, compression ratios, and AUC trade-offs should be verified against the original PDF.

Context in the Field

Neural recommendation and retrieval has evolved from matrix factorization and deep CTR models, through sequence Transformers, toward LLM-based and generative recommendation (Gen-Rec). Core constraints remain unchanged: sparse user behavior, massive item catalogs, and multi-objective business trade-offs. Embedding compression techniques (quantization, hashing, low-rank factorization, mixed precision) address the memory/latency side of this triangle while preserving accuracy.

Related Entries in This Collection

The Scandinavian Embedding Benchmarks
Arctic-Embed 2.0: Multilingual Retrieval Without Compromise (arXiv:2412.04506)
BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity (arXiv:2402.03216)
BGE-en-ICL / BGE-ICL: Making Text Embedders Few-Shot Learners (arXiv:2409.15700)
Beyond Benchmarks: Evaluating Embedding Model Similarity for Retrieval (arXiv:2407.08275)
BitNet Text Embeddings (arXiv)

Takeaways for Practitioners

Researchers: Reproduce the core comparisons and check whether the paper reports statistical significance and compute costs.
Engineers: Evaluate pluggable compression modules against your existing CTR stack; consider index update frequency and embedding-version compatibility.
Product teams: Track user-perceivable effects (latency, memory cost savings) rather than only offline AUC deltas.

References

Original paper: arXiv:2502.15355 — <https://arxiv.org/abs/2502.15355>