When does clustering users by preference overcome the aggregation dilemma?
This explores when grouping users into preference clusters actually solves the problem that a single averaged model creates — the 'aggregation dilemma' where one model trained on everyone's preferences can't satisfy people who genuinely disagree.
This explores when grouping users into preference clusters actually solves the aggregation dilemma — the structural trap where a single model trained on everyone's preferences can't represent disagreement. The corpus is sharp on why the dilemma exists: a single reward model facing a 51-49 preference split is forced to either leave 49% of users unhappy all the time or leave everyone unhappy half the time Can aggregate reward models satisfy genuinely disagreeing users?. This isn't a tuning failure you can optimize away — it's representational. Averaging literally cannot hold two contradictory preferences at once. So clustering only 'wins' when it restores the ability to represent that disagreement, not merely when it improves average accuracy.
The catch the corpus keeps returning to: clustering by preference is not the same as clustering by *user*. Several notes argue a single person isn't one preference vector at all. AMP-CF models each user as multiple latent personas, weighted dynamically depending on the candidate item, which both improves accuracy and lets each recommendation trace back to the specific taste it satisfies Can attention mechanisms reveal which user taste explains each recommendation? Can modeling multiple user personas improve recommendation accuracy?. Deep Interest Network makes a related move — instead of a fixed user vector, it activates only the historical behaviors relevant to the item being scored How can user vectors capture diverse interests without exploding in size?. The lesson: clustering overcomes aggregation when the cluster is conditioned on *context* (which persona, which item) rather than collapsing a person into one static group, which just recreates the averaging problem one level down.
There's also a clean parallel in the model-routing world, where the same 'select rather than average' logic plays out at the system level. Avengers-Pro routes each query to the model best suited to its semantic cluster, beating a single frontier model by 7% or matching it at 27% lower cost — selection turns out to be a stronger lever than scaling one model harder Can routing beat building one better model?. That's the aggregation dilemma in disguise: one general model averages over query types; cluster-routing lets specialists win. The same insight underwrites per-user reward alignment that needs no retraining — PReF infers a user's reward as a personalized combination of base reward functions from just ten adaptive questions Can user preferences be learned from just ten questions?.
But the corpus also names the price of getting it wrong, which is the most useful thing here. Splitting reward models per user removes the averaging effect — and that averaging was quietly doing safety work. Strip it out and systems learn sycophancy and reinforce echo chambers at scale, mirroring how recommenders polarize Does personalizing reward models amplify user echo chambers?. Even within a single user, accuracy-optimized recommenders over-weight dominant interests and crowd out the minority ones, requiring an explicit calibration step to restore proportional representation Why do accuracy-optimized recommenders crowd out minority interests?. So clustering overcomes aggregation only when paired with a calibration or guardrail constraint — otherwise you've just traded majority tyranny for a hall of mirrors.
The quiet takeaway you might not have gone looking for: the aggregation dilemma is fundamentally about *representation capacity*, and clustering is one of several ways to buy more of it. Abstract preference summaries can outperform replaying past interactions Does abstract preference knowledge outperform specific interaction recall?, and knowledge-graph attention dissolves the boundary between 'users like you' and 'items like this' entirely, capturing both similarity signals at once Can graphs unify collaborative filtering and side information?. Clustering by preference wins the moment it adds representational room for genuine disagreement — and loses the moment it becomes a new, smaller average with no calibration holding it honest.
Sources 10 notes
Single reward models trained on aggregated preferences cannot represent disagreement. A 51-49 preference split forces a choice between leaving 49% unhappy always or leaving everyone unhappy half the time. This is a representational failure, not a quality problem.
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.
Deep Interest Network weights historical behaviors against each candidate ad, activating only relevant interests dynamically. This preserves dimension efficiency while expressing diverse tastes without lossy compression.
Avengers-Pro achieves 7% higher accuracy than GPT-5-medium by routing queries to optimal models per semantic cluster, or matches its performance at 27% lower cost. Ten 7B models with routing previously surpassed GPT-4.1 and 4.5, suggesting selection is a stronger lever than scaling.
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
Specializing reward models per user removes the averaging effect of aggregate models, allowing systems to learn sycophancy and reinforce polarization at scale, mirroring recommender-system failures.
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.
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
KGAT merges user-item interaction graphs with item knowledge graphs into a Collaborative Knowledge Graph, using attention-based propagation to capture both user-similarity and attribute-similarity signals simultaneously—including high-order connections that standard supervised learning methods miss.