Mixed Dimension Embeddings with Application to Memory-Efficient Recommendation Systems

2019/09/25 by Antonio Ginart, Maxim Naumov, Ginart, Antonio +7 · 8 citations
Computer Science · #Advanced Graph Neural Networks #Caching and Content Delivery #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Recommender Systems and Techniques

paper · pdf · doi:10.48550/arxiv.1909.11810

openalex publication_date 2019/09/25 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Embedding representations power machine intelligence in many applications, including recommendation systems, but they are space intensive -- potentially occupying hundreds of gigabytes in large-scale settings. To help manage this outsized memory consumption, we explore mixed dimension embeddings, an embedding layer architecture in which a particular embedding vector's dimension scales with its query frequency. Through theoretical analysis and systematic experiments, we demonstrate that using mixed dimensions can drastically reduce the memory usage, while maintaining and even improving the ML performance. Empirically, we show that the proposed mixed dimension layers improve accuracy by 0.1% using half as many parameters or maintain it using 16X fewer parameters for click-through rate prediction task on the Criteo Kaggle dataset.

Citations

• Matrix Factorization Techniques for Recommender Systems
• A Simpler Approach to Matrix Completion
• Deep Learning Recommendation Model for Personalization and Recommendation Systems
• DeepFM: A Factorization-Machine based Neural Network for CTR Prediction
• Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism

Cited by

• Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
• Dimension Mask Layer: Optimizing Embedding Efficiency for Scalable ID-based Models
• Understanding Embedding Scaling in Collaborative Filtering
• Random Offset Block Embedding Array (ROBE) for CriteoTB Benchmark MLPerf\n DLRM Model : 1000\× Compression and 3.1\× Faster Inference
• ECRM: Efficient Fault Tolerance for Recommendation Model Training via Erasure Coding
• Differentiable Neural Input Search for Recommender Systems
• Memory-efficient Embedding for Recommendations
• Learnable Embedding Sizes for Recommender Systems

Related

• Learnable Embedding Sizes for Recommender Systems
• Fine-Grained Embedding Dimension Optimization During Training for Recommender Systems
• Mixed-Precision Embeddings for Large-Scale Recommendation Models
• Mixed-Precision Embedding Using a Cache
• Dimension Mask Layer: Optimizing Embedding Efficiency for Scalable ID-based Models