Prototypical Extreme Multi-label Classification with a Dynamic Margin Loss

Date

April 2025

Source

KDD Main Research Track

Authors

Kunal Dahiya
Diego Ortego
David Jiménez

Summary of Our Research

In this work, we designed an efficient algorithm for recommendation engines that scales up to millions of items. This innovation not only applies to all our product matching use cases but is also a key component for LLMs (RAG systems). Our thorough evaluations on public datasets from Amazon and Wikipedia have shown that our algorithm outperforms those from big tech labs like Google and Microsoft Research, using significantly fewer resources (1 vs 16 GPUs).