Embedding Items at Scale: Comparing GNN-Based and ID-Based Item Embeddings in the Yandex Ecosystem

Transformer-based sequential recommenders depend heavily on the item embedding strategy: item embeddings can either be pretrained separately or learned end-to-end with the transformer. This case study compares pretrained industrial graph neural network item embeddings with end-to-end trainable item embeddings from both a cost and a quality perspective, across two mature production recommender systems at Yandex (Yandex Market and Yandex Music) and a low-resource dataset sampled from Yandex Lavka production logs, for which the data and code are publicly available. A separate pretraining stage helps when training data is limited, but provides no worthwhile benefit for large-scale models trained on extensive datasets.