Can heterophily-based social recommendations reduce opinion polarization?
This explores whether recommending content through friends who *disagree* with you — rather than friends who share your tastes (homophily) — could counteract the opinion convergence and echo chambers that recommender systems tend to produce.
This reads the question as two linked claims: (1) that social recommendation built on *different* tastes works at all, and (2) that doing so might push back against polarization. The corpus has strong material on the first and suggestive — but not direct — material on the second.
On the first claim, the collection is unusually clear. Social Poisson Factorization shows that friends with *different* preferences actually outperform recommenders that assume your friends share your taste Can friends with different tastes improve recommendations?. The mechanism is the interesting part: networks add value not through taste similarity but through *influence on anomalous choices* — your contrarian friend is precisely the one who introduces you to the thing you'd never have found inside your own preference bubble. So heterophily isn't just tolerable in recommendation; it's a feature that homophily-based methods leave on the table.
On the second claim, the corpus describes the polarization machine that heterophily would have to fight. Recommendation feeds aren't neutral plumbing — they're persuasion infrastructure where network topology itself drives opinion convergence at population scale How do recommendation feeds shape what people see and believe?. And the *type* of recommender matters: frequently-bought-together vs. co-viewed networks produce measurably different convergence-vs-divergence patterns, because each pulls together different audience segments with different priors Do different recommender types shape opinion convergence differently?. That's the lateral payoff — it implies the convergence/divergence dial is real and *tunable by design choice*, which is exactly what a heterophily intervention would exploit.
The corpus also shows the counter-current that makes this hard. The strongest analogue to polarization-by-recommendation here is personalized reward models: specializing models per user removes the averaging effect of an aggregate model and lets the system learn sycophancy and reinforce polarization — explicitly mirroring recommender-system failures Does personalizing reward models amplify user echo chambers?. And accuracy-optimized recommenders systematically crowd out minority interests, requiring post-hoc reranking just to restore proportional representation Why do accuracy-optimized recommenders crowd out minority interests?. The drift toward homogeneity isn't a bug you forgot to fix — it's what pure accuracy optimization *converges to*.
So the honest synthesis: nothing here runs the direct experiment ('inject heterophilous recommendations, measure polarization drop'). But the pieces line up into a plausible mechanism — heterophilous social signals demonstrably surface choices outside your bubble Can friends with different tastes improve recommendations?, convergence is a topology-and-design property rather than a fixed law Do different recommender types shape opinion convergence differently?, and the default optimization target actively fights you Does personalizing reward models amplify user echo chambers?. The thing you didn't know you wanted to know: in this corpus, diversity and *accuracy* aren't opposites you trade off — the friends-with-different-tastes method got better recommendations *because* of the difference, which means de-polarizing and improving the feed might be the same move, not competing ones.
Sources 5 notes
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.
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.
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.
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.