Practical sequential recommender walkthrough

Recommendation software usually guesses from what similar users liked. Sequential recommenders instead read a user’s clicks like a watch history and predict the next item in order. (arxiv.org) A new code-first walkthrough from MLWhiz uses that idea to explain three common model families: GRU4Rec, SASRec, and BERT4Rec. The thread ties each model to PyTorch examples and a Steam games dataset rather than only paper diagrams. (x.com) (kaggle.com) GRU4Rec comes from a 2015 paper that applied recurrent neural networks to session-based recommendation, where the model reads one action after another and updates a hidden state. SASRec, published in 2018, replaced that recurrence with self-attention so the model can weigh earlier items in a sequence when predicting the next one. (arxiv.org 1) (arxiv.org 2) BERT4Rec, published in 2019, adapted the masked-token training style from BERT to recommendation by hiding items in a sequence and asking the model to recover them from left and right context. That differs from next-item-only training and gives practitioners a second transformer baseline to compare against SASRec. (arxiv.org 1) (arxiv.org 2) The Steam examples matter because game activity is naturally sequential: a player buys, launches, and reviews titles over time. One widely reused processed Steam dataset contains 200,000 interactions from 12,393 users across 5,155 games, while MLWhiz’s notebook describes a larger raw setup with about 88,000 users and 32,000 games. (github.com) (kaggle.com) The production angle is the part many tutorials skip. MLWhiz’s notebook includes deployment patterns, and the paper trail behind these models shows why teams care: GRU4Rec was built for short sessions, SASRec was pitched as more efficient than comparable recurrent and convolutional models, and BERT4Rec added a different training objective that many later benchmarks still test. (kaggle.com) (arxiv.org 1) (arxiv.org 2) That framing now sits next to a newer line of work that treats recommendation more like sequence generation at very large scale. Meta researchers’ 2024 HSTU paper described “generative recommenders” and reported gains of up to 65.8% in NDCG on synthetic and public datasets, along with 5.3x to 15.2x speedups over FlashAttention2-based transformers on 8,192-length sequences. (arxiv.org) Meta followed that work with ULTRA-HSTU in February 2026, describing additional model-and-system changes for large-scale sequential recommendation. That puts older names like GRU4Rec, SASRec, and BERT4Rec in a clearer lineage: recurrent models, attention models, masked-sequence models, then generative recommender systems tuned for long histories and industrial traffic. (arxiv.org) (arxiv.org) The practical value of a walkthrough like this is that it gives engineers the vocabulary used in interviews and design docs without hiding the implementation. In recommendation, the jump from “users who liked this also liked” to “predict the next action in a sequence” is the shift that explains why these model names keep showing up in production discussions. (substack.com) (arxiv.org)