Yandex Study: ID Embeddings Outperform GNN Embeddings in Large-Scale Recommenders

_reachsumit · x · 2026-07-30

Yandex published a case study on arXiv comparing item embedding strategies in large-scale recommendation systems, specifically contrasting pretrained GNN-based embeddings with end-to-end trainable ID embeddings.

The research reveals that while a separate pretraining stage benefits low-resource settings with limited training data, it provides no worthwhile benefit for large-scale models trained on extensive datasets, where ID embeddings emerge as the winner.

Original post →

More from Research

Research channel →