How do influence and homophily differ as mechanisms in social networks?
This explores the difference between two ways networks shape behavior — influence (your friends change what you do) versus homophily (you already resemble the people you're connected to) — and why telling them apart matters for systems built on social data.
This explores the difference between two ways networks shape behavior: influence, where connections actually change what you do, and homophily, where you simply resemble the people you're already linked to. The distinction sounds academic, but the corpus shows it has teeth — systems that confuse the two end up recommending the wrong things to the wrong people. The sharpest treatment comes from work on Social Poisson Factorization Can friends with different tastes improve recommendations?, which argues that most social recommenders quietly assume homophily — they pull friends' representations together on the theory that connected people share tastes. But that work finds the real signal lives elsewhere: friends with *different* preferences are what improve recommendations, because the network adds value through influence on your anomalous, out-of-character choices, not by confirming your usual taste. Homophily predicts what you'd have picked anyway; influence is what moves you somewhere new.
Once you separate the two, you can watch influence operate as its own force over time. In partner-selection games Do humans learn to prefer AI partners over time?, people start out biased against AI agents but gradually come to prefer them through repeated interaction — preference here is *built* by experience, not inherited from similarity. That's influence as a learning process rather than a static resemblance. The same dynamic shows up in adversarial form in multi-agent systems How does workflow position shape attack propagation in multi-agent systems?, where malicious signals travel farthest when injected at high-influence positions — influence concentrates wherever dependencies converge, which is precisely the kind of structural leverage homophily can't explain.
The corpus also shows how homophily and influence get entangled at population scale, which is what makes them so easy to mistake for each other. Research framing recommendation feeds as persuasion infrastructure How do recommendation feeds shape what people see and believe? points out that network topology itself drives opinion convergence — feeds don't just reflect who's similar, they actively push connected people toward agreement. And work on opinion convergence by recommender type Do different recommender types shape opinion convergence differently? surfaces the mechanism cleanly: 'frequently-bought-together' and 'co-viewed' networks produce different convergence patterns because each attracts a different audience segment with different priors. That's homophily (who shows up) and influence (how being shown things together shifts ratings) operating in the same system — the observed convergence is a blend, and you can only untangle it if you know which knob you're turning.
The quiet payoff: if you treat resemblance as influence, you build systems that amplify what people already are — echo chambers, popularity loops, convergence toward the mean. If you can isolate genuine influence, the network becomes a source of productive surprise — the friend whose different taste introduces you to something you'd never have found alone. The interesting questions in social systems usually live in that gap, where the network does something to you rather than just describing you.
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.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.
FLOWSTEER demonstrates that malicious signals propagate farther when injected into high-influence subtasks, and that framing them as evidence rather than instruction causes downstream agents to relay them. Influence concentrates where dependencies converge, making position-aware attacks far more effective.
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.