Can recommender systems correct for ratings that have been socially shaped?
This explores whether recommender systems can detect and correct for ratings that aren't independent signals of quality — ratings that have been bent by prior ratings, social influence, and rater noise — rather than treating every star as an honest preference.
This explores whether recommenders can correct for ratings that aren't clean signals — stars shaped by what other people already said, by the rater's mood that day, or by the feed that put the item in front of them. The first thing the corpus does is dismantle the premise that ratings are independent in the first place. Moe and Trusov decomposed online ratings into baseline quality, social-dynamics influence, and noise, and found prior ratings meaningfully bend the ones that follow — with effects that compound forward through future ratings Do online ratings actually reflect independent customer opinions?. On top of that social drift sits plain inconsistency: the same person rates the same item differently across sessions, sometimes by multiple stars, so a rating encodes rating-behavior as much as preference Why do the same users rate items differently each time?. So 'correcting' has two distinct problems hiding inside it — social contamination and individual noise.
Here's the uncomfortable part the corpus surfaces: the recommender isn't a neutral observer that could clean this up — it's often the thing doing the shaping. Recommendation networks like 'frequently bought together' versus 'co-viewed' drive connected products toward convergent or divergent opinions depending on which audience each one funnels in Do different recommender types shape opinion convergence differently?. Stepping back further, feeds behave as persuasion infrastructure: feed weights move producer behavior and network topology pushes opinions to converge, with rating contamination as one of the compounding channels How do recommendation feeds shape what people see and believe?. A system that generates the social shaping can't straightforwardly subtract it.
Where the corpus gets constructive is in techniques that don't try to recover a 'true' rating at all, but instead correct for the distortions structurally. Accuracy-optimized models over-weight whatever interest dominates the data — exactly the crowding you'd expect when popular opinion snowballs — and post-hoc reranking can re-impose proportional representation without retraining Why do accuracy-optimized recommenders crowd out minority interests?. The popularity bias has a quieter root cause too: when embedding dimensions are too small, the model overfits toward popular items and locks in long-term unfairness, which means the fix is a design hyperparameter, not a post-hoc patch Does embedding dimensionality secretly drive popularity bias in recommenders?. Different layer, same disease — social shaping shows up as popularity concentration, and you can attack it at the loss, the architecture, or the reranking stage.
A second strategy is to route around contaminated rating signals by leaning on structure that's harder to socially distort. Social Poisson Factorization deliberately recommends through friends with *different* tastes rather than similar ones, extracting value from influence on anomalous choices instead of from herd-like taste similarity Can friends with different tastes improve recommendations?. And modeling a user as multiple attention-weighted personas rather than one averaged vector lets the system surface minority interests that a socially-averaged rating would bury, while explaining which persona each recommendation serves Can modeling multiple user personas improve recommendation accuracy? Can attention mechanisms reveal which user taste explains each recommendation?.
The honest synthesis: the corpus has no method that reads a rating and backs out the social influence baked into it — the contamination compounds forward and the recommender is part of the loop. What it offers instead is correction by design — calibration constraints, fairness-aware dimensionality, diverse-friend signals, and multi-persona modeling that refuse to let dominant opinion crowd out the rest. The reframe worth taking away: the fixable problem isn't 'what would this rating have been without social pressure,' it's 'how do we stop the system from amplifying that pressure downstream.'
Sources 9 notes
Moe and Trusov decomposed ratings into baseline quality, social-dynamics influence, and error, finding that prior ratings meaningfully affect subsequent ones. These effects have both immediate sales impact and long-term compounding effects through future ratings, though high opinion variance can eventually dampen the distortion.
Amatriain et al. found that the same user gives substantially different ratings to the same item across sessions, shifting by multiple stars. This noise stems from temporal inconsistency, rater-specific biases, and anchoring effects—making ratings reflect both preference and rating-behavior rather than stable preference alone.
Research shows that frequently-bought-together and co-viewed recommendation networks produce different opinion convergence patterns. The mechanism: each recommender type attracts different audience segments with different prior expectations, shaping both who sees products together and how they rate them.
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.
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.
Social Poisson Factorization uses friends' diverse tastes to recommend items outside users' usual preferences, outperforming methods that pull friends' representations together. Networks add value through influence on anomalous choices, not taste similarity.
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.