A Probabilistic Model for Using Social Networks in Personalized Item Recommendation

“We first review previous research on using social networks to help recommend items to users. A crucial component of SPF is that it infers the influence that users have with each other. In previous work, some systems assume that user influence (sometimes called “trust”) is observed [27]. However, trust information beyond a binary yes/no is onerous for users to input, and thus observing trust beyond “following” or “friending” is impractical in a large system. Others assume that trust is propagated [2] or computed from the structure of the network [10]. This is limited in that it ignores user activity, which can reveal the trust of a user for some parts of the network over others; SPF captures this idea. Information diffusion [8, 12] also relies on user activity to describe influence, but focuses on understanding the widespread flow of information. A final alternative is to compute trust from rating similarities between users [9]. However, performing this computation in advance of fitting the model confounds general preference similarity with instances of influence—two people with the same preferences might read the same books in isolation.

Other research has included social information directly into various collaborative filtering methods. Ref. [36] incorporates the network into pairwise ranking methods. Their approach is interesting, but one-class ranking methods are not as interpretable as factorization, which is important in many applications of recommender systems [15]. Refs. [25, 28, 34] have explored how traditional factorization methods can exploit network connections. For example, many of these models factorize both user-item data and the user-user network. This brings the latent preferences of connected users closer to each other, reflecting that friends have similar tastes. Refs [24, 35] incorporate this idea more directly by including friends’ latent representations in computing recommendations made for a user.

Our model has a fundamentally different approach to using the network to form recommendations. It seeks to find friends with different preferences to help recommend items to a user that are outside of her usual taste. For example, imagine that a user likes an item simply because many of her friends liked it too, but that it falls squarely outside of her usual preferences. Models that adjust their friends’ overall preferences according to the social network do not allow the possibility that the user may still enjoy this anomalous item. As we show in Section 3, using the social network in this way performs better than these previous approaches.”