Modern Recommender Architectures

The boundary. Ranking & Recommendation System Architecture covers the funnel and its compute budget, classical candidate generation (matrix factorization, item-to-item, two-tower with in-batch negatives, ANN indexes), basic pointwise/pairwise/listwise rankers, the value formula in one paragraph, position bias, cold start, re-ranking and the offline-online gap. This lesson does not repeat any of that. It covers the model architectures that appeared after about 2016: explicit feature-interaction networks, multi-task towers with gating, sequence encoders over user histories, generative recommenders that replace the funnel with a sequence model, and where LLMs fit.

Neighbouring lessons own these topics, and this one links to them instead of repeating them:

• Embedding mechanics (InfoNCE, the logQ derivation, hashing-trick basics): Embeddings & Representation Learning.
• Bandit exploration: Contextual Bandits for Personalization. Estimating a new policy from logs: Off-Policy Evaluation & Offline RL.
• Long-run objectives and delayed rewards: Long-Term Value & Delayed Reward Systems. Interference in marketplace experiments: Advanced Experimentation: Interference, Marketplaces & Platforms.
• Graph-based retrieval: see Graph ML & Graph Neural Networks.
• The three end-to-end cases: Case Study: Video Recommendation (YouTube / Netflix-style Homepage), Case Study: Ad Click-Through Rate Prediction (Meta / Google Ads-style) and Case Study: Long-Term Engagement Recommender.

How to read the claims.

• Everything marked verified was run with the scripts in /tmp/rec/ (numpy, scikit-learn and torch 2.14 CPU, one machine, small synthetic data) while this lesson was written. Those numbers illustrate mechanisms and are not benchmarks.
• Statements about papers (DCN-v2 2021, MMoE 2018, PLE 2020, SASRec 2018, BERT4Rec 2019, DIN 2018, TIGER 2023, HSTU 2024 and others) are from memory, hedged, and contain no quoted result numbers. Check the papers before citing them.
• Anything that could not be run here (HSTU scaling, semantic-ID quality, LLM reranking) is explained conceptually and labeled not measured.

Interviewers ask the same four questions of every modern recommender:

1. What does the model see (features and history)?
2. What does it predict (one task or many, and how are they combined)?
3. How is it trained and kept fresh?
4. How do you know it is better?

The sections below follow that order.