Predictive Optimisation

What it is

Two claims about phones are usually made in the same breath, and on the face of it they cannot both be true.

The first is that a feed is a slot machine. Refresh it and the reward arrives on an unpredictable schedule, the arrangement Ferster and Skinner found produced the most persistent behaviour of anything they tested. The second is that the feed knows you, that it holds a model of your attention built from years of your own behaviour and serves what that model predicts you will stay for.

Randomness and prediction are opposite engineering goals. A slot machine holds no model of the person pulling the handle, and its payout ratio is fixed before anybody sits down. A ranking system does nothing but hold a model of you, and every engineering hour spent on it goes into making the next item less of a surprise. If a recommender were improved to the point of perfect accuracy, the schedule would disappear entirely, because nothing would ever fail to land.

Predictive optimisation is the name for that second mechanism taken on its own terms. A candidate set of items is scored against a model of the individual, the scores are ranked, and the top of the ranking is served. The objective function is predicted engagement, recalculated per person, and the system improves by being more right. Nothing in it is stochastic in the way a payout table is stochastic. The randomness a user experiences is not designed in; it is what is left over after the model has done its best.

That leftover is the whole argument. The model works better than chance and nowhere near certainty. What that produces at the user's end is a stream in which most items are mildly worth continuing and an unpredictable minority are genuinely rewarding. The ratio between those two categories is set by the accuracy of the model rather than by a randomiser. From the seat of the person holding the phone, it is a variable payout arriving after an unpredictable number of swipes, which is to say it is a variable-ratio schedule that nobody designed, tuned to one person rather than to an average customer.

This page is filed as not-a-research-claim and the label is doing real work. No primary source is cited for the claim that recommender accuracy generates a variable-ratio schedule. It is an argument from two mechanisms, presented as such, and the experiment that would separate it from the designed-schedule account has not been run.

In effect

Research: the experiment nobody has run

The honest research entry on this page is an absence, and it is worth being precise about its shape. Separating the two accounts would require degrading a live recommendation system for a randomised subset of real users, making it measurably less accurate while changing nothing else, and then measuring what happened to session length. If accuracy is the active ingredient, engagement falls. If the schedule is, engagement may hold or even rise, since a worse model produces a lumpier stream. The only organisations that could run that experiment are the ones with the least reason to publish the result, and the reason is mundane rather than sinister. Until somebody does, both accounts remain live.

Everyday life: the stretch of feed that was fine

The experience the model predicts is not excitement and not boredom. It is a run of items that are each mildly worth continuing, punctuated by something that genuinely lands, with no way of telling in advance which swipe produces which. That is exactly what a good but imperfect model would deliver, and it is exactly the profile Ferster and Skinner described for a variable ratio. The tell is the same one in both cases: a run of unrewarding items carries no information, because unrewarding runs are ordinary, so nothing in the experience ever signals that it is time to stop.

Commerce: the metric that cannot settle it

Published engagement figures are the numbers everybody has and they are the wrong instrument. They report how much people use a product and how that changes when a ranking model is improved, which is consistent with both accounts and discriminates between neither. A rise in session length after a ranking upgrade is what you would predict if accuracy were the active ingredient. It is equally what you would predict if a better-tuned model happened to produce a better-tuned schedule. Every organisation with the data is measuring the one quantity that cannot answer the question, and has no commercial reason to notice.

What it does not say

It does not say that recommendation systems are designed to be addictive. The argument on this page runs the other way. It says the schedule can arrive as a by-product of an accuracy objective, with nobody having to intend it, which is a claim about mechanisms and not about motives. Establishing intent requires evidence about intent, and this is not that.

It does not settle the argument against the slot-machine account. The strongest version of that case is that the distinction between designing a schedule and stumbling into one is a distinction about intent, and intent is not what does the damage. Voinea, Marin and Vica's 2024 paper in Topoi makes the structural case without needing any account of origin, and the machine gambling literature documents the design intent in slot machines in detail. That position may simply be right.

It is not a term of art from the recommender-systems literature. Predictive optimisation is this publication's own construct, coined for one article to hold two mechanisms apart. Nobody else uses it, and a reader looking for it in a machine learning textbook will not find it.

It makes no quantitative claim. No accuracy figure, no ratio, no effect size appears on this page, because none has been established for the mechanism described here.

Who developed it

Nobody. This is the article's own construct rather than a term taken from a source, and it is recorded that way in the frontmatter, with originator unknown and confidence none. The two mechanisms it holds apart do have origins: the schedule is Ferster and Skinner's, catalogued in 1957, and the ranking machinery is ordinary industrial practice with no single author.

---