Ranking and Objectives: What Should the Model Optimise?

The ranking stage receives a few hundred candidates and must order them. Under the hood, orderings come from predictions: for each candidate, the model estimates the probability or magnitude of some user response, and the list is sorted by a function of those estimates.

That sentence hides the entire difficulty in two words: some response. Which response?

• Predict clicks, and the system learns to prize whatever gets clicked, including regret-clicks on sensational packaging.
• Predict completion, and short items win a structural advantage.
• Predict watch or listen time, and length gets rewarded independent of quality.
• Predict explicit likes, and you optimise for the tiny, unusual minority who presses buttons.

Key idea: the objective is not a technical detail; it is the product's values, executed at scale, millions of times a day. Every recommender heuristic you have ever disliked, clickbait rising, feeds full of one topic, is usually not a model failing its objective but a model succeeding at an objective that was chosen badly.

So before any architecture: what signal is worth predicting, and what does each signal secretly reward? That is this lesson.

Recommenders train almost entirely on implicit feedback, behaviour observed in passing, because explicit feedback is scarce: most users never rate, and the ones who do are systematically unusual.

The menu, with each signal's built-in slant:

Two structural facts about this table run the field. First, every signal is a proxy: the thing you want, the user was glad this was recommended, is unobservable, and each proxy diverges from it somewhere specific. Second, the divergences are exploitable, and models find the exploits reliably: optimise clicks hard enough and the catalogue's suppliers learn to manufacture clickability, an ecosystem response no offline evaluation predicts.

The craft is therefore not finding the one true signal, it does not exist, but combining proxies so their failure modes partially cancel, which is where multi-objective ranking comes from.

Before any objective can be optimised, there is a problem with the training data itself, and it is the same demon that haunts search evaluation: the log records what users did with what they were shown, and what they were shown was chosen by the previous model.

Position bias is the sharpest form. Items placed at the top get seen and clicked more, regardless of quality; items below the fold barely get examined at all. A model trained naively on this log learns two things at once, what users like, and what the old system used to rank highly, inseparably blended. The second part is pure contamination: it teaches the new model to imitate the old one's placements, including its mistakes.

Production rankers predict several responses at once, a multi-task model with one head per signal, and combine the predictions into a single ranking score:

where each head's prediction, click probability, expected watch time, like probability, skip probability with a negative weight, is scaled by a weight expressing how much the product values that outcome.

Two honest facts about the weights. First, they are a product decision wearing a mathematical costume: the vector of weights is the executive answer to "what do we want more of", and no gradient descent chooses it for you. Teams tune weights through online experiments, watching not one metric but the whole dashboard, because raising the click weight moves watch time, complaints, and creator behaviour all at once.

Second, the combination is where the proxy failures get their partial cancellation: watch time's length bias is checked by completion's brevity bias; click's packaging bias is checked by the skip head punishing fast regret. The weights are less a formula than a system of checks, and reading a company's ranking weights, where they have been published or leaked, reads like reading its values, which is exactly what they are.

In practice: expect the score function to also include hand-built terms no model predicted, freshness boosts, quality floors from human review, creator-diversity bonuses, because some product values have no behavioural signal to learn from.

The deepest objective problem is temporal. Every signal in the table is measured within minutes of the recommendation; the outcomes the business actually cares about, retention, subscription renewal, the user's sense that the product respects their time, unfold over months and cannot be attributed to a single ranked item.

The tension is not hypothetical. Optimising short-term engagement can degrade the long-term outcome through mechanisms that are well understood qualitatively: sensational content that wins the click and erodes trust; binge patterns that maximise tonight's minutes and tomorrow's uninstall; monocultures that deepen engagement per session while narrowing the reasons to return.

What serious teams do about a gap that cannot be closed directly:

• Guardrail experiments: ranking changes ship behind long-running holdbacks, small user groups kept on the old system for months, so retention effects invisible in a two-week test have somewhere to show up.
• Long-horizon proxies in the objective: heads that predict return-tomorrow or session-after-this-one, weak signals, but pointed at the right target.
• Constraint-based formulations: maximise engagement subject to floors on diversity, freshness, or quality metrics, encoding as constraints what cannot be learned as objectives.
• Periodic value audits: reading the top of the funnel, what actually got recommended this week, against stated product values, because the aggregate dashboard hides composition.

None of this is fully solved, and the honest position is the useful one: the objective function is a permanent negotiation, not a settled artefact, and the teams that treat it as finished are the ones whose products drift somewhere nobody chose.

One technical property becomes load-bearing the moment scores are combined or thresholded: calibration. A model is calibrated when its predicted probabilities match observed frequencies, among items given a 10 percent click probability, about 10 percent actually get clicked.

Ranking alone does not need it: any monotone distortion of the scores produces the same order. But the multi-objective sum does need it, because the weighted combination adds predictions from different heads, and if the click head runs systematically hot while the watch-time head runs cold, the weights stop meaning what the product team thinks they mean: the effective policy silently differs from the intended one. Ads systems need it even more desperately, since predicted click probability multiplies directly into auction prices.

Miscalibration creeps in through routine causes: training on down-sampled negatives, distribution shift between training and serving, and the feedback data itself moving as the ranking changes. The standard hygiene: measure calibration continuously per head, segment by the slices that matter, new items versus established, head versus tail users, and re-calibrate with simple post-hoc maps, Platt scaling or isotonic regression, rather than retraining the world.

It is unglamorous machinery, but it is the difference between a weighted objective that encodes the product's chosen trade-offs and one that encodes them plus a hidden, drifting error term.

Close the lesson the way a real team closes a quarter: by interrogating a live objective. A video platform ranks with 0.3 times click probability, plus 0.6 times predicted watch minutes, capped per item, plus 0.4 times like probability, minus 0.5 times early-skip probability.

What is this objective buying, and where will it leak?

The heavy watch-time weight, even capped, tilts toward longer items; the cap limits single-item domination but does nothing across the session, so bingeable series content still compounds. The click term is small, sensible, packaging bias is real, but combined with the skip penalty it creates a specific incentive: content that is clicked and held, which mostly means strong openings; expect creators to discover this within months and front-load accordingly. The like term quietly over-serves the expressive minority's tastes to everyone. Nothing in the sum values diversity, freshness, or return probability: those must live in constraints and retrieval mix, or they live nowhere.

That analysis, objective to incentives to predictable ecosystem response, is the transferable skill, and it needs no access to the model: published weights, or even observed behaviour, support the same reading.

One loose end remains, and it is the biggest one. This whole lesson trained on logs of what the system previously chose to show, and treated that as given. What happens when today's model decides tomorrow's training data, in a loop, is the final lesson: feedback loops, and the exploration that keeps a recommender from eating its own tail.