Can confidence levels improve recommendations compared to single-number ratings?
This reads the question as: does it help to model uncertainty and richer signal — probabilities, attention weights, proportions — rather than collapsing a recommendation down to one predicted rating number? The corpus has a lot to say here, mostly under different vocabulary.
This reads the question as asking whether modeling *how confident* a system is — or more broadly, capturing richer signal than a single predicted score — beats the classic "predict one rating number per item" approach. Several papers in the collection answer yes, though almost none use the word "confidence." The shared insight is that a single number throws away structure the system actually needs to rank well.
The sharpest version is about probability rather than point estimates. When a recommender predicts one rating per item independently, it never forces items to compete for limited attention. Switching to a multinomial likelihood — where items share a probability budget — directly aligns training with what you actually want, a ranked top-N list, and beats the Gaussian and logistic alternatives that treat each rating as its own isolated number Why does multinomial likelihood work better for ranking recommendations?. So "confidence" expressed as competing probabilities outperforms "confidence" expressed as a lone score.
Another angle: a single rating assumes a user has one monolithic taste. AMP-CF instead splits each user into multiple personas and uses attention weights to decide which persona is confident about a given candidate item — improving accuracy *and* explaining why each item was recommended, without a separate diversity step Can modeling multiple user personas improve recommendation accuracy? Can attention mechanisms reveal which user taste explains each recommendation?. Here the "confidence levels" are per-persona weights, and they carry information a flat score can't.
There's also a cautionary thread about what happens when you optimize a single relevance number too hard. Ranking purely by per-item score crowds out a user's secondary interests, so calibration — preserving the *proportions* of what someone likes — restores diversity without hurting accuracy Do accuracy-optimized recommendations preserve user interest diversity?. And squeezing representations too small quietly bends the whole system toward popular items, an unfairness that compounds over time Does embedding dimensionality secretly drive popularity bias in recommenders?. Both suggest a single confidence number, naively maximized, can actively mislead.
The surprising counterpoint is from a different domain: more numbers don't automatically mean more trustworthy. Users prefer answers with more citations even when those citations are irrelevant — citation count becomes a decoupled trust heuristic rather than a real signal Do users trust citations more when there are simply more of them?. The lesson for recommendation is that richer-looking signal only helps if it's *grounded*; the win comes from confidence that reflects genuine competition, proportion, or persona structure — not from decorating a recommendation with extra numbers.
Sources 6 notes
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.
AMP-CF separates user representation into latent personas weighted by attention to the candidate item. This candidate-conditional approach improves accuracy by adapting the user representation at prediction time and produces inherent explanations for why items were recommended.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.
Steck's research shows that ranking by per-item relevance naturally produces lists dominated by a user's primary interest, even when they have documented secondary interests. Enforcing calibration via post-hoc reranking restores proportional representation without sacrificing overall accuracy.
Research shows that when user/item embedding dimensions are too small, recommender systems overfit toward popular items to maximize ranking quality. This compounds over time as niche items receive insufficient exposure, and cannot be fixed post-hoc without treating dimensionality as a fairness hyperparameter.
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.