Recommendation systems platforms explained with examples

The recommendation problem is not one problem. It changes shape with what is being recommended, and the shape matters more than the company. A system that recommends thirty-second clips and one that recommends jobs share a name and almost nothing else: different item lifetimes, different costs of a wrong guess, different evidence available. Sort products by shape and their behavior becomes predictable.

No platform is named here. Each row below describes a category several products occupy.

What to take away

• Six properties of the thing being recommended decide most of a recommender's behavior: how long items live, how much it costs to try one, whether repeats are welcome, how large the catalog is, what explicit signals exist, and what a bad recommendation costs.
• Read a product's row and you know what its system must optimize for, what it will get wrong, and what a publisher into that product should care about.
• Advice written for one row is wrong for the others in specific, predictable ways.

The six properties

Product type	Item lifetime	Cost to try one	Repeats welcome	Catalog	Explicit signals	Cost of a bad pick
Short clips	Hours to days	Seconds	Rarely	Enormous, fast-growing	Almost none; skips do the work	Nearly zero; the next one is a swipe away
Long video	Weeks to years	Many minutes	Sometimes	Large	Some: subscriptions, ratings	Moderate; a wasted half hour
Music and audio	Years	Minutes	Constantly	Large, slow-growing	Playlists, saves, skips	Low; but repetition fatigue accumulates
Text posts	Hours	Seconds	Never	Enormous	Follows, replies	Low individually; high in aggregate for mood and trust
Shopping	Months	Money	Sometimes	Large	Purchases, returns, searches	High; money and trust
Jobs and listings	Weeks	Hours of effort	Never	Moderate	Applications, filters, saves	High for both sides; a wasted application
News	Hours	A minute	Never	Moderate, fast-turnover	Follows, subscriptions	High for trust; low per item
People to follow	Indefinite	A tap	Never	Everyone	The graph itself	Low, but errors compound the graph

Where the last row applies and the graph itself is the evidence, the shape of the graph is the product: clusters form, and a recommendation is mostly a bet about which cluster somebody belongs to.

What each shape forces

Short lifetime pushes evidence collection to the first hours. Where items die in a day, the system cannot wait for evidence to accumulate. It has to guess from attributes and from the first handful of responses, which means early behavior is everything and late quality is invisible. A publisher into a short-lifetime product should think about the first minute and nothing else.

Low cost to try makes exploration cheap. Where a wrong pick costs the viewer a swipe, the system can afford to show uncertain items often, because the penalty is small and the information is valuable. Where a wrong pick costs money or an application, exploration is expensive and the system leans hard on what it already knows, which is bad news for anything new.

Repeats change the objective. Music is the odd row: the best recommendation is often something the person has heard a hundred times. A system that treats "already consumed" as a reason not to show something is wrong for music and right for news. This one property explains why the same person can love one product's recommendations and hate another's built on similar machinery.

Catalog size decides whether the first pass matters. With a moderate catalog, a careful model can consider most of it. With an enormous one, the cheap first pass that shortlists candidates decides most outcomes before the careful model runs, and the constraints that force a feed into stages apply in full.

Explicit signals reduce guessing. A purchase, an application, a save: each says more than a view, because the producers of evidence differ in how much intent they can carry. Products rich in explicit signals can afford to be conservative. Products with almost none, short clips above all, have to infer everything from what people did not do, and skips become the loudest signal in the system.

The cost of a bad pick sets the tolerance for surprise. Where being wrong is free, the system can be adventurous, and the pleasant accidents that make a discovery product feel alive, the quality sometimes called serendipity, are affordable. Where being wrong costs money or trust, surprise is a liability and the recommendations become safe, repetitive, and dull, on purpose.

The general problem and its standard approaches are laid out in the reference entry on recommender systems; what the table adds is that the same approaches produce different behavior depending on which row they are pointed at.

Reading the table as a publisher

• Short clips and text posts: the first hours decide. Prepare the opening, publish when your audience is present, and do not judge an item after its window has closed. The engines behind a spike matter here more than anywhere.
• Long video and audio: completion and return matter more than the opening. Slow accumulation is normal; an item can be discovered a year on.
• Shopping and listings: matching matters more than appeal. Be accurately described, in the words people search with, and the system can do the rest. Being exciting is nearly irrelevant.
• News: freshness dominates; the item is over before evidence forms. Reputation of the source carries more weight than any single item.
• People: the graph is the product. Every recommendation of you is based on who already connected to you, so the early connections shape all the later ones.

Common questions

Why do short-clip products feel more responsive than long-video ones?

Cheap tries and short lifetimes let them collect evidence fast and act on it fast. The responsiveness is a property of the item shape, not of better engineering.

Why does a shopping recommender keep showing me what I already bought?

Because for some purchases a repeat is the right recommendation, and the system cannot always tell which kind yours was. It is guessing the "repeats welcome" column for one item, and guessing wrong.

Is a news recommender just a text-post recommender with sources?

Structurally close, with one change: the cost of a bad pick includes the reader's trust, which makes source reputation an input in a way it is not for casual posts.

Which row is hardest to build for?

Jobs and listings. Two-sided cost, no repeats, expensive tries, and a moderate catalog where every wrong match wastes someone's hours. The two families of evidence are both thin there.