Recommendation architectures

An architecture describes how a recommender learns and scores items. A trained model version is the result of fitting that architecture to a particular dataset. Start in Models, choose eligible candidates, then compare them on the same training snapshot before deployment. There is no universally best architecture. Catalogue size, history length, available signals, latency and the objective you measure all affect the choice. The console’s readiness checks determine whether your data meets the requirements for a particular training run.

Neighbourhood and content models

Recommender	How it works	Main tradeoff
ItemKNN	Finds items that co-occur in user histories and scores neighbours of items a user engaged with	A useful behavioural baseline; new items without co-occurrence evidence need another source of relevance
VS-KNN	Matches the current session to similar past sessions	Useful for session-driven intent; depends on relevant session history rather than a durable user profile
Content TF-IDF	Matches item text to the content in the user’s history	Can score new items from descriptions; limited descriptions reduce its ability to distinguish them
EASEᴿ	Learns linear item-to-item relationships from interaction histories	A compact modelling approach whose memory requirements grow rapidly with catalogue size; respect the catalogue’s size limit
iALS	Learns latent user and item factors from implicit feedback	Learns broad preference patterns; relies on interaction coverage and meaningful feedback weights
LightGCN	Propagates information across the user–item interaction graph	Can find indirect relationships; sparse evidence and popularity effects still need evaluation

Sequence models

These models use the order of a user’s interactions. Choose them when changes in intent or the next likely action matter, and when your events contain reliable ordering and enough history.

Recommender	How it works
GRU4Rec	A recurrent network summarizes the ordered interaction sequence to predict what comes next
SASRec	A causal transformer attends to earlier items in the sequence
BERT4Rec	A bidirectional transformer learns by reconstructing masked items in histories
HSTU (Generative Recommenders)	A sequential architecture uses position and time information to model changing interaction histories

Longer history or a more complex model is not automatically better. Compare sparse-history and dense-history users separately, and keep time-based evaluation consistent between candidates.

Retrieval followed by ranking

A retrieval stage selects candidates; a ranking stage orders that smaller set. The ranker can improve the order of retrieved candidates, but cannot recover an item the retriever never supplied.

Recommender	Retrieval and ranking approach
ItemKNN + MLP ranker	Neighbourhood candidates reordered by a neural network using retrieval, history and content features
ItemKNN + Multi-task ranker	Neighbourhood candidates with separate output heads for available engagement signals, combined using chosen weights
VS-KNN + GBDT	Session-neighbour candidates reordered by gradient-boosted trees
EASEᴿ + LambdaMART	Linear item-to-item candidates reordered by a listwise boosted ranker; inherits EASE’s catalogue-size limit
Two-Tower + DCN-v2	Separate user and item encoders retrieve candidates, followed by a deep-and-cross ranking network
Two-Tower Token-Mixer	Two-tower retrieval followed by a token-mixing ranking architecture
ItemKNN + Option-Attention reranker	Neighbourhood retrieval followed by candidate-specific attention over the user’s recent history
Cheap full pass + Option-Attention reranker	A low-cost catalogue scoring pass selects candidates for a more detailed attention-based pass

The catalogue’s trained ranking stage is distinct from the optional stages you configure in a context’s Pipeline. Check both when investigating the final order.

Option scoring and generative models

Recommender	What distinguishes it
Option-Attention Scorer (full catalogue)	Scores each item against its own attended view of the user’s history; cost grows with the number of items scored
Calibrated Multi-Outcome Option Scorer	Predicts separate engagement outcomes and calibrates those predictions on held-out data before combining them
TIGER	Encodes items as semantic codes and generates the codes of a likely next item
Semantic-ID Language Model	Uses semantic item tokens and a staged training curriculum for generative retrieval
Multi-Token Foundation Recommender	Predicts multiple future interactions from a unified history representation that also uses item content
Preference-Aligned Session Generator	Generates a recommendation sequence jointly, then aligns it with a reward objective
Unified Action Transformer	Uses a related transformer design for retrieval and ranking with semantic item representations

A generative recommender predicts items or item codes; it does not necessarily generate new product descriptions or chat responses. Its catalogue coverage, decoding limits and latency still need to be measured. Some catalogue entries implement published architectural ideas at a smaller scale. They are not the original company’s model or a promise of that company’s results. Read each entry’s confidence label and caveat: an official implementation, a reference, a reproduction and a proxy provide different levels of correspondence to the named research.

Non-personalised comparison points

Popularity ranks items by total interaction count. Recency-Weighted Popularity gives recent interactions more influence. Both return an order that does not depend on the requesting user’s preferences. Entries marked Baseline are comparison references, not selectable training candidates. These comparisons matter: a model that cannot improve on a simple baseline for your objective may not justify added training or serving cost.

Choose and evaluate

Check data readiness and select a small set of plausible architectures.
Use the same signal template and snapshot when comparing runs.
Inspect relevance, cold-start performance, long-tail exposure, novelty, diversity and multi-objective behaviour. Check how many datasets or users contributed to a comparison.
Review latency and cost alongside quality. Catalogue reference figures are comparison estimates, not a guarantee for your workload.
Evaluate on your own held-out events, then use a controlled experiment to assess live behaviour.

The console catalogue is the current source for availability, requirements and reference results. Use Model Training for run metrics and Models for deployment and rollback.

Overview

Relevance

Data

Insights

Models

Developers

Account

Recommendation architectures

Neighbourhood and content models

Sequence models

Retrieval followed by ranking

Option scoring and generative models

Non-personalised comparison points

Choose and evaluate

​Neighbourhood and content models

​Sequence models

​Retrieval followed by ranking

​Option scoring and generative models

​Non-personalised comparison points

​Choose and evaluate

Neighbourhood and content models

Sequence models

Retrieval followed by ranking

Option scoring and generative models

Non-personalised comparison points

Choose and evaluate