Should recommender objectives optimize for individual item relevance or list-level coverage?
This explores a real tension in recommender design: train each item to be maximally relevant on its own, or judge the whole list by how well it covers the spread of a user's tastes — and the corpus suggests the second matters more than per-item scoring admits.
This question reads as a choice between two objectives: rank each item by how well it matches the user (item relevance), or shape the whole list so it represents the user's full range of interests (list-level coverage). The corpus comes down fairly hard on one side — pure item relevance has a built-in failure mode, and several notes show it from different angles.
The cleanest statement of the problem is in the calibration work. When you rank purely by per-item relevance, the list naturally fills up with a user's single dominant interest, even when their history clearly documents secondary ones — the minority tastes get crowded out not because they're wrong, but because each individual item from the dominant interest scores slightly higher Do accuracy-optimized recommendations preserve user interest diversity?. The fix proposed there is a post-hoc reranking pass that enforces proportional representation, and the striking part is that it restores coverage *without* sacrificing accuracy Why do accuracy-optimized recommenders crowd out minority interests?. That's the key tell: the two objectives aren't actually a tradeoff at the list level — optimizing item relevance alone just leaves coverage on the table.
What makes this more than a niche concern is that the same blind spot shows up in the model's plumbing, not just its ranking rule. When embedding dimensions are too small, the system overfits toward popular items to squeeze out ranking quality, and niche items get starved of exposure in a way that compounds over time and can't be patched after the fact Does embedding dimensionality secretly drive popularity bias in recommenders?. So 'optimize item relevance' quietly becomes 'optimize for the popular and the dominant' at multiple layers — the loss function, the embedding capacity, the ranking step — all pushing the same direction.
The more interesting answer the corpus offers is that you may not have to choose between a relevance model and a separate diversity step at all. The multi-persona approach represents a user as several latent tastes, weighted by attention to each candidate item, so coverage falls out of the representation itself: each recommendation traces back to the specific persona it satisfies, and the usual post-hoc diversity reranking becomes unnecessary Can attention mechanisms reveal which user taste explains each recommendation?Can modeling multiple user personas improve recommendation accuracy?. That reframes the whole question — coverage isn't a constraint you bolt onto a relevance objective, it's what you get when you stop modeling the user as a single point.
Worth knowing: even the choice of training objective leans this way. Switching a model's likelihood to multinomial — which forces items to compete for probability mass — beats alternatives precisely because that competition aligns training with top-N *list* ranking rather than scoring items in isolation Why does multinomial likelihood work better for ranking recommendations?. So the answer isn't 'relevance or coverage.' It's that relevance scored item-by-item is the weaker framing, and the strongest systems bake list-level structure — competition, proportional taste, multiple personas — into the objective from the start.
Sources 6 notes
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.
Accuracy-optimized models systematically miscalibrate by over-weighting dominant user interests. A post-processing reranking algorithm that enforces calibration constraints can restore proportional representation without retraining the underlying model.
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.
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.
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.
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.