The Invisible Funnel: How Recommenders Find What Matters

In a feed with thousands of articles and videos, your phone has milliseconds to surface the few things worth your attention. Candidate generation is the invisible first filter that narrows chaos to choice before anything gets ranked. This episode unpacks matrix factorization, embeddings, and the matchmaking metaphor behind how recommender systems find needles in haystacks at billion-scale speed. 00:00:00 - The Funnel Explained: Candidate Generation and the Millisecond Problem 00:01:42 - The Mechanism: Matrix Factorization and Latent Spaces 00:03:18 - Profiles as Embeddings: The Go-Between 00:05:26 - Latent Factor Space: The Takeaway 00:07:09 - The Omiai Metaphor: How Profile Preparation Works 00:08:53 - Batch Processing: Why Negatives Are Free 00:10:50 - The Upper Layers and Speed 00:12:25 - Scaling to Billions: Graphs at Global Scale 00:14:10 - Union Operations and Coverage 00:15:51 - The Denominator Problem 00:17:43 - Open Questions: How Many Candidates Is Enough? --- Sources & further reading: • Koren, Bell, Volinsky, Matrix Factorization Techniques for Recommender Systems, IEEE • Yi et al., *Sampling-Bias-Corrected Neural Modeling for Large Corpus Item • Covington, Adams, Sargin, Deep Neural Networks for YouTube Recommendations, RecSys '16 • Malkov, Yashunin, *Efficient and robust approximate nearest neighbor search using • Douze et al., The Faiss library, arXiv:2401.08281, 2024-01-16 (v4 2025-10-23) • Johnson, Douze, Jégou, Billion-scale similarity search with GPUs, arXiv:1702.08734 • Medvedev, Wu, Gordon, "Powered by AI: Instagram's Explore recommender system", Meta AI • All URLs fetched on 2026-10-15. • Computer*, August 2009: https://www.cs.columbia.edu/~blei/fogm/2023F/readings/KorenBellVolinsky2009.pdf • (course-hosted copy; IEEE Xplore record not fetched) — CF vs content, cold start, inner products • SGD/ALS, biases, implicit feedback, Netflix Prize RMSE figures. • Recommendations*, RecSys '19, September 2019, doi 10.1145/3298689.3346996 • two towers, in-batch: https://storage.googleapis.com/gweb-research2023-media/pubtools/5716.pdf • negatives, logQ correction, multiple nominators, index pipeline, Recall@K, live A/B. • sampled negatives, nearest-neighbour serving, blending sources, the precision/recall wording. • Hierarchical Navigable Small World graphs*, arXiv:1603.09320, 2016 (v4 2018) • layers, skip-list analogy, ef and M, recall definition.: https://arxiv.org/abs/1603.09320 • brute force vs IVF, nprobe, MIPS, recall at shortlist size.: https://arxiv.org/abs/2401.08281 • 2017-02-28 — — the original FAISS GPU paper.: https://arxiv.org/abs/1702.08734 • blog, 2019-11-25: https://ai.meta.com/blog/powered-by-ai-instagrams-explore-recommender-system/ • seed accounts, KNN lookup, candidate sources, 500 → 150 → 50 → 25. This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.