The engine behind your feed
Sessions 2 and 3 built the machinery: neural nets that learn from examples, and embeddings that place things in a meaning-space. Recommendations are the first place you see that machinery earning its keep every single day. A recommender does not "understand" you - it notices that your behavior rhymes with other people's, and borrows their taste to fill in yours. That one trick powers a startling share of what you watch, buy, and hear. This session shows you how it works, where it breaks, and the single question a leader must keep asking: what are we optimizing for?
Two ways to recommend 5 min live
There are two fundamentally different ways to guess what you will like. One looks at people; the other looks at items. Most products you use run both at once and blend the answers - but keeping them apart in your head is the key to understanding any recommender.
LiveCollaborative vs content-based, and why blend3 min▶
- Collaborative filtering uses the crowd's behavior: "people who rated things the way you did also liked X." It never needs to know what an item is about - only who liked what. This is the engine's heart.
- Content-based uses the item's own features: "you liked this sci-fi space thriller, here is another sci-fi space thriller." It leans on tags and embeddings (Session 3), not on other users.
- Each has a blind spot. Collaborative cannot help a brand-new item nobody has touched yet; content-based only ever echoes what you already like and rarely surprises you.
- So real systems blend both - and add popularity, recency, and business rules on top. When you hear "the algorithm", it is almost always a weighted mix, not one clean idea.
Spotify's Discover Weekly is the classic blend: collaborative filtering finds listeners whose playlists overlap yours and surfaces what they play that you have not heard, while content-style audio features keep the picks in a coherent mood. The magic feeling - "how did it know?" - is just your taste-twins, at scale.
Self-studyMatrix factorization in one paragraph2 min read▶
The famous engine under collaborative filtering is matrix factorization. Picture the giant grid of users-by-items, mostly empty (nobody rates everything). Factorization finds a small set of hidden "taste factors" - unnamed dials like "gritty vs cozy" or "mainstream vs cult" - such that every user is a blend of factors and every item is a blend of factors, and multiplying the two back together reproduces the ratings you do have and fills in the ones you do not. Nobody labels those factors; the math discovers them from the grid. It is the same move as embeddings (Session 3): compress messy behavior into a few meaningful numbers, then measure distance in that space.
The engines you use daily + cold start 4 min live
These are not products with recommendations bolted on - they are recommenders wearing a product. But every one of them faces the same awkward first moment: a brand-new user or a brand-new item, with no history to learn from. That is the cold-start problem, and how a company solves it shapes your very first experience.
LiveRecommender-first products, and the cold-start problem3 min▶
- Netflix ranks and reshuffles the whole home screen per person; Amazon's "customers also bought" drives a huge slice of the cart; Spotify builds playlists from listening overlap; TikTok's entire For You feed is a recommender with almost no other UI. The recommendation is the product.
- Cold start is the catch: with zero history, collaborative filtering has nothing to compute. A new user has no taste-twins yet; a new item has no one who has rated it yet.
- The fixes are all "borrow signal from somewhere else": ask a few onboarding questions ("pick 3 shows you like"), lean on content features (genre, tags) until behavior arrives, and fall back to popularity ("here is what is trending") so the feed is never empty.
- Why leaders care: the first-session experience decides whether a new user ever comes back. Cold start is not an edge case - it is every user's day one.
The 2006 Netflix Prize put $1M on improving recommendations by 10% - that is how much the engine was worth to them. And roughly 35% of Amazon's sales are attributed to its recommendation system. Recommendations are not a feature you add for delight; for these companies they are the business model.
When recommenders go wrong 5 min live
A recommender optimizes for whatever metric you point it at - and it is relentless. Point it at "time on screen" and it will faithfully learn that outrage and cliffhangers keep you scrolling. The failures below are not bugs; they are the system doing exactly what it was told, which is why the human choice of objective matters more than the math.
LiveBubbles, runaway loops, and the objective that matters3 min▶
- Filter bubbles: click a few of one kind of thing and the feed narrows to more of it. Helpful at first, then quietly cuts off everything else - you stop seeing what you never clicked.
- Engagement optimization gone bad: if the metric is time-on-screen, the system learns that outrage, fear, and clickbait win. It is not malicious - it is a faithful optimizer pointed at the wrong number.
- Feedback loops: the model trains on the outputs it produced yesterday, so its own choices become tomorrow's "truth". Small early biases amplify into big ones.
- Why a human-set objective matters: the fix is not better math, it is a better target - optimize for long-term satisfaction, diversity, or "did you finish and feel good", not just the next click. Someone has to choose that.
Self-studyImplicit vs explicit signals, and designing for serendipity2 min read▶
Recommenders learn from two kinds of signal. Explicit signals are deliberate - a star rating, a thumbs-up, a "not interested". Implicit signals are behavioral - a click, a hover, seconds watched, a re-play. Implicit data is abundant but noisy: a click is not a like (you might have clicked and immediately regretted it, or lingered because you were confused). Systems that treat every click as endorsement are the ones that spiral into clickbait. The mature response is to add diversity and serendipity as deliberate design choices - reserve slots for the unexpected, cap how narrow the feed can get, and weight explicit "I loved this" over raw clicks. Good recommendation is as much product judgment as it is modeling.
Rate a few, watch it predict the rest ★ 10 min · everyone plays
Collaborative filtering, live and in your hands. Rate a handful of items and the engine finds the phantom user whose taste most matches yours, then borrows their ratings to predict yours. Watch the recommendations move as you change your mind.
Rate 3 or 4 items honestly. Read the recommendations it produces and the "because your taste-twin..." explanation underneath.
Now change one rating - flip a high score to low. Watch the recommendations shift as your taste-twin changes. The engine has no opinions; it only mirrors the crowd.
Say it out loud: no one ever described these items to the machine. It knows nothing about genre or content - only who liked what. That is collaborative filtering in one breath.
Design the cold start ★ 6 min · everyone designs
Pick a product you know well (yours, or one you use). Imagine a brand-new user opening it for the very first time - the feed is empty and the engine knows nothing.
Design the first session. What 3 questions or signals would you collect to bootstrap recommendations? Onboarding picks? Location? Which content features could stand in until behavior arrives?
Decide your popularity fallback: while you wait for signal, what safe, broadly-liked items keep the screen from being empty?
Share one design. Notice how fast "just recommend things" turns into real product decisions about what to ask and when.
Spot the objective ★ 6 min · everyone diagnoses
Four real feeds. For each, name the metric it most likely optimizes and the failure mode that metric risks - then propose a healthier objective.
A news feed. What is it optimizing - clicks? time? What failure mode does that invite (outrage, doom-scrolling, clickbait headlines)?
A short-video feed. Likely watch-time and completion. What behavior would that reward if pushed to the extreme? What bubble forms?
A shopping feed and a music feed. Do the same for each: guess the metric, name the risk, then propose a healthier objective (satisfaction, diversity, "came back next week").
Compare answers. The same technology serves all four - the objective, chosen by a human, is what makes one helpful and another harmful.
Several large platforms have publicly re-tuned their feeds away from raw engagement - down-weighting content that keeps people scrolling but leaves them feeling worse, and adding "are you still enjoying this?" style checks. That is a business admitting the metric was the problem, not the model. The literate leader asks the objective question before the harm, not after.
This week ◐ 25 min total
- Name the strategy: for three recommendations you get this week (a show, a product, a song), guess whether it came from people-similarity (collaborative) or item-similarity (content-based) - and why.
- Finish your cold-start design from Apply-along 2 - write the 3 onboarding signals and the popularity fallback for a product you care about.
- Watch your own bubble: notice one feed narrowing on you this week. What did you click that started it? Would a diversity slot have helped?
- Read the self-study cards above: matrix factorization, and implicit vs explicit signals.
Three questions before you go 🎯 ◐ 90 seconds
1 · What is the core difference between collaborative filtering and content-based recommendation?
Collaborative filtering matches you to similar users and borrows their taste; content-based matches items by their features. Real systems blend both.
2 · What is the cold-start problem?
No history means no taste-twins and no ratings. You borrow signal from elsewhere until behavior arrives - and it happens to every user on day one.
3 · Where do filter bubbles come from?
Bubbles are the engine faithfully doing its job: reward the click, show more of it, narrow the feed. The fix is a better objective and deliberate diversity, not better accuracy.
Concepts covered & where they come from
This course teaches the working intuition of each idea and credits the canon. Leaders get judgment; practitioners get the Self-study lane and the names to go deeper.