Recsperts - Recommender Systems Experts

In episode 29 of Recsperts, I welcome Craig Macdonald, Professor of Information Retrieval at the University of Glasgow, and Aleksandr “Sasha” Petrov, PhD researcher and former applied scientist at Amazon. Together, we dive deep into sequential recommender systems and the growing role of transformer models such as SASRec and BERT4Rec.

Our conversation begins with their influential replicability study of BERT4Rec, which revealed inconsistencies in reported results and highlighted the importance of training objectives over architecture tweaks. From there, Craig and Sasha guide us through their award-winning research on making transformers for sequential recommendation with large corpora both more effective and more efficient. We discuss how recency sampling (RSS) reduces training times dramatically, and how gSASRec overcomes the problem of overconfidence in models trained with negative sampling. By generalizing the sigmoid function (gBCE), they were able to reconcile cross-entropy–based optimization results with negative sampling, matching the effectiveness of softmax approaches while keeping training scalable for large corpora.

We also explore RecJPQ, their recent work on joint product quantization for item embeddings. This approach makes transformer-based sequential recommenders substantially faster at inference and far more memory-efficient for embeddings—while sometimes even improving effectiveness thanks to regularization effects. Towards the end, Craig and Sasha share their perspective on generative approaches like GPTRec, the promises and limits of large language models in recommendation, and what challenges remain for the future of sequential recommender systems.

Enjoy this enriching episode of RECSPERTS – Recommender Systems Experts.

Don’t forget to follow the podcast and please leave a review.