How does choosing fatigue affect which ranking positions matter most to users?
This explores position bias — the fact that users overwhelmingly act on whatever sits at the top of a ranked list, somewhat regardless of relevance — and whether the corpus frames that concentrated attention as a kind of decision-economizing on the user's part.
This reads your question as being about position bias: when scanning a ranked list is effortful, people don't evaluate everything — they act on the top few items, and the rest barely register. The corpus doesn't study "choosing fatigue" as a psychology experiment, but it treats the engineering symptom directly. YouTube's multi-objective ranker Why do ranking systems need to model selection bias explicitly? builds a dedicated "position tower" whose entire job is to absorb the fact that an item gets clicked partly because it was placed high, not because it was the best match. If you don't subtract that out, the model learns the position, not the preference.
The reason this matters more than it first appears: position bias isn't just a measurement nuisance, it's a feedback loop. Because users economize their attention on top slots, those slots collect the clicks, the clicks become training data, and the model concludes those items deserved the top slots — a self-fulfilling loop the same note describes as a "degenerate equilibrium" that amplifies a system's own past decisions. The recommendation-feeds note How do recommendation feeds shape what people see and believe? traces where this leads at population scale: selection biases and rating contamination compound until the feed isn't reflecting preferences so much as manufacturing them. So which positions "matter most" is partly an artifact the system has to actively unlearn.
The sharpest lateral angle is the distinction between learning and choosing Can utility-weighted training loss actually harm model performance?. If users only ever act on the top of the list, you're tempted to weight training toward getting that decision right — utility-weighted loss. The surprising finding is that this backfires: optimizing directly for the choice degrades the model's underlying feature learning, and you do better by learning with a symmetric loss and adjusting for the decision afterward. In other words, designing a ranker around the few positions users actually engage with can quietly starve the model of the signal it needs to rank well in the first place.
This is also why the choice of likelihood function turns out to matter Why does multinomial likelihood work better for ranking recommendations?: a multinomial likelihood forces items to compete for a shared probability budget, which aligns training with top-N ranking — exactly the regime where user attention is concentrated and only the top handful get a real look. The competition between items mirrors the competition for the user's limited willingness to scan.
And there's a cost to chasing those top positions in real time. Netflix's in-session work How can real-time recommendations stay responsive and reproducible? shows you can improve ranking ~6% by re-ranking as a session unfolds — moving the right item up before the user's patience runs out — but only by paying in latency, call volume, and bugs you can't reproduce. The thing worth taking away: "which positions matter" isn't a fixed property of the list. It's set by how much effort the user is willing to spend, and the entire stack — debiasing towers, loss functions, likelihoods, real-time re-ranking — is bent around that limited budget of attention.
Sources 5 notes
YouTube's multi-objective ranker uses MMoE for conflicting objectives and a shallow position tower to remove selection bias from training data. Without both mechanisms, models converge on degenerate equilibria that amplify their own past decisions.
Research shows recommendation systems operate as political actors: feed weights influence producer behavior, network topology drives opinion convergence, and automation enables targeted persuasion at population scale. These effects compound through rating contamination and selection biases.
Asymmetric loss functions correctly incentivize choosing but degrade representation learning by reducing gradient signals for substantive feature acquisition. Training with symmetric loss then adjusting predictions post-hoc outperforms direct utility-weighted training on the same utility objective.
Liang et al. show that switching VAE likelihoods from Gaussian/logistic to multinomial achieves state-of-the-art results because enforced probability competition between items directly aligns training with top-N ranking objectives. Rebalancing KL regularization further improves performance.
Netflix's in-session adaptation improves ranking by 6% relative, but precomputing is impossible when signals arrive mid-session. This forces runtime recomputation, increasing call volume, timeout risk, and making bugs harder to reproduce.