Authors: Gleb Mezentsev, Danil Gusak, Ivan Oseledets, Evgeny Frolov

Published on: September 27, 2024

Impact Score: 7.8

Arxiv code: Arxiv:2409.18721

Summary

• What is new: A novel Scalable Cross-Entropy (SCE) loss function for recommender systems that is more efficient in terms of time and memory usage.
• Why this is important: High computational overload and excessive GPU memory usage in modern recommender systems when dealing with large item catalogs.
• What the research proposes: Introducing SCE loss function that approximates CE loss using a selective GPU-efficient computation strategy without compromising recommendation quality.
• Results: SCE significantly reduces peak memory usage by up to 100 times while retaining or even improving recommendation performance metrics.

Technical Details

Technological frameworks used: Selective GPU-efficient computation strategy

Models used: SCE loss function approximates softmax distribution through maximum inner product search

Data used: Multiple datasets with large-size catalogs

Potential Impact

Companies operating large recommender systems and those developing large-scale models, such as large language models, could benefit significantly from these insights.

Want to implement this idea in a business?

We have generated a startup concept here: EffiRecom.