GRU4Rec
Most traffic on a shop or a news site is anonymous. There is no profile to look up and no history to factorise — just the four or five things clicked since the page loaded. GRU4Rec treats that short burst as the whole input and predicts what comes next.
The model
A gated recurrent unit consumes the session's item events in order, carrying the state of the session in its hidden vector, and emits a score for candidate items at each step. There is no user embedding anywhere in the architecture — identity is replaced by sequence, which is what makes the model usable for a first-time visitor.
The engineering that made it work
Two choices matter as much as the recurrence. Sessions range from two clicks to hundreds, so training uses session-parallel mini-batches: many sessions advance side by side, each replaced by a fresh one as it ends. And a softmax over a full catalogue is unaffordable, so negatives are sampled — popularity-biased and shared across the mini-batch — under pairwise ranking losses such as BPR and TOP1.
Why it matters
GRU4Rec is the result that made sequential recommendation credible, beating well-tuned item-kNN baselines on session data, and the self-attentive models that followed — SASRec, BERT4Rec — kept its framing and swapped out the recurrence. Ignore order and a session collapses into a bag of items: the shopper who has just moved from phones to laptop sleeves keeps being shown phones.
GRU4Rec is a session-based recommender that runs a gated recurrent unit over the items clicked so far in a session and predicts the next one, with no user embedding at all. It made deep sequential recommendation practical through session-parallel mini-batching and sampled-negative ranking losses in place of a full-catalogue softmax. It is the model for anonymous traffic and short-lived intent, and the direct ancestor of SASRec and BERT4Rec.
Walk into a shop as a complete stranger and the assistant still has something to go on: the three things you have picked up and put down in the last two minutes.
GRU4Rec is a recommender built to work exactly that way. It ignores who you are — there may be no account at all — and reads only the short trail of clicks in your current visit, in order, to guess what you will want to see next. The order is the point: someone who has moved from looking at phones to looking at phone cases has changed what they are shopping for, and a system that merely counted the clicks would miss it.