Can friends with different tastes improve recommendations?

The conventional approach to incorporating social networks in recommendation is to assume friends have similar tastes (homophily) and pull connected users' latent representations together via regularization. Each user's preferences are shaped to be closer to their friends' preferences. But this confounds two different things: people with similar tastes happen to be friends (homophily) versus people influence their friends' specific choices (influence). The first is just preference-similarity-by-correlation. The second is causal — your friend's recommendation made you read this book.

Social Poisson Factorization (SPF) decouples them. The model uses friends with different preferences to help recommend items outside the user's usual taste. Imagine a user who likes an item simply because many of her friends liked it, even though it falls outside her usual preferences. Models that pull friends' overall preferences together would miss this — they assume tastes converge, so they discount the anomalous item. SPF allows the network to surface specifically anomalous-but-influence-driven items.

The empirical claim is that this approach outperforms previous network-aware factorization methods. The conceptual claim is that "trust" or "social regularization" methods misidentify the channel through which networks help: they treat the network as a way to enforce taste similarity, but the actual value is in finding items the user wouldn't reach through their own taste alone — items their friends' diverse tastes expose them to.

Source: Recommenders Personalized