Learning-Based vs. Innate Friend Selection: Two Explanations

Two Explanations of Homophily

The fact that people tend to befriend those who are similar to them — homophily — is one of the most robust findings in network science. But there are two competing explanations for the underlying mechanism.

Preference-based homophily assumes that people derive greater utility from interactions with similar others. Similar friends share interests, values, and life contexts, making the relationship intrinsically more valuable (Currarini, Jackson & Pin, 2009). In this model, the expected reward of an interaction is higher when the partners are similar.

Learning-Based Theory

Learning-based homophily proposes an alternative mechanism (Gitmez & Zarate, 2022): the expected value of an interaction is equally high for similar and dissimilar pairs. The difference lies in the speed of learning. Similar individuals share communication styles, common knowledge, and mental models (Mayhew et al., 1995; Carley & Palmquist, 1992), allowing them to discover faster whether a relationship is valuable. In the formal model’s “exploration phase” — drawing on exponential bandit theory (Keller et al., 2005) — interaction partners first explore the value of their relationship. Similar pairs receive a positive signal faster and sustain the interaction more often. Dissimilar pairs need more time for the same insight and frequently abandon the exploration before a positive signal arrives.

Crucially, the two models make different empirical predictions: in the preference model, proximity strengthens the existing homophily effect — similar pairs benefit more from proximity. In the learning model, by contrast, proximity acts as a substitute for similarity: it extends the exploration phase and helps dissimilar pairs in particular discover the value of their relationship. Experimental data from Peruvian boarding schools support the learning model.

Preference-Based Theory

Gitmez and Zarate (2022) formalize the friendship process as a decision model with uncertain information, based on the bandit theory of Keller et al. (2005). Their model generates three testable predictions: (1) friendship patterns show homophily, (2) proximity fosters friendships, and (3) the proximity effect is stronger for dissimilar pairs. The preference model of Currarini et al. (2009) predicts that prediction (3) goes in the opposite direction. The randomized experiment at COAR boarding schools in Peru (N ≈ 5,500) confirms all three predictions of the learning model. Kets and Sandroni (2019) offer a related information-theoretic approach: homophily arises because similar individuals receive correlated action signals and overcome strategic uncertainty more easily — a mechanism that also relies on information advantages rather than preference differences.

The two models are not mutually exclusive — in reality, both learning and preference effects are likely at play. Gitmez and Zarate (2022) formally show that with sufficiently strong preference differences (a strong “extensive margin”), the learning model result is reversed. The evidence from Peru suggests that the learning effect dominates, but this need not hold universally. The assumption of uniform priors (all interactions are equally valuable in expectation) is a strong simplification. Moreover, the model does not capture dynamic network processes such as transitivity and reciprocity, which additionally influence friendship formation in real networks. Gallen and Wasserman (2022) also show that homophily in mentor choice disappears once information about mentors is provided — a finding that likewise supports the information-theoretic explanation.

What the Peru Study Shows

The current state of research on this aspect is summarized below.

Why It Matters

The current state of research on this aspect is summarized below.

Frequently Asked Questions

Do we choose friends out of preference or habit?

The tendency to befriend similar people — homophily — is one of the most robust findings in network science. But there are two competing explanations for the underlying mechanism.

What is learning-based homophily?

Learning-based homophily proposes an alternative mechanism (Gitmez & Zarate, 2022): the expected value of an interaction is equally high for similar and dissimilar pairs. The difference lies in the speed of learning.

Which explanation is correct?

The two models are not mutually exclusive — in reality, both learning and preference effects are likely at play.

Can you influence friend selection?

Preference-based homophily assumes that people derive greater utility from interactions with similar others. Similar friends share interests, values, and life contexts, making the relationship intrinsically more valuable (Currarini, Jackson & Pin, 2009).