What if there were a smarter way to expand your social network? That is to say, what if you were offered a reason to follow someone beyond “people you may know” on Flickr, Facebook, and LinkedIn, “recommended blogs” on Tumblr, and “who to follow” on Twitter, for example?
The recommendation of users to other users is one of the most fundamental and important features of online social-networking platforms. This capability helps people get a faster start in building their networks, which as a result drives engagement and loyalty. Given that growing the user-base and maintaining a high level of engagement are key factors for the success (or death) of these billion-dollar businesses, the importance of user-recommendation systems is clear.
These systems usually exploit the structure of a given social graph in order to predict which new connections, or social links, are likely to appear in the future: This task is known as “link prediction” in machine learning literature. As an example, the most basic form of a rule for link prediction is “triadic closure”: If person A follows person B, and person B follows person C, then it is likely that A might be interested in following C as well. Of course, much more sophisticated techniques and algorithms are behind the real-world systems we mention above. In fact, a social-networking platform maintains much more information than just a social graph (who is connected to whom). The system may also know which school you frequented, the places you live and work, and the music and movies you like. While this wealth of information is surely utilized to recommend connections, it is not capitalized upon to explain a recommendation.
Enhancing recommendations with explanations adds an important layer of trust for someone using a social network. And when someone has a more compelling explanation for a suggestion, they are more likely to click “follow” and expand their network. While this premise is well understood in classic collaborative-filtering recommender systems, providing explanations in the context of the user-recommendation systems we describe is still largely underdeveloped. In fact, in most real-world systems, the explanations given for user recommendations are something along the lines of, “You should follow person Z because your contacts X and Y do the same.”
Based on this consideration, in our recent research paper at the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), we introduce “Link Prediction with Explanation,” a novel machine learning task for which we devised a model dubbed, “Who to Follow and Why.” Our foundational observation is rooted in sociology, where it is known as “common identity and common bond theory.” In simple terms, people create connections either due to common interests (topical links) or based upon common social settings such as family, work, school, etc. (social links). Our framework not only recommends new social connections, but for each suggested link it decides whether it is topical or social, and depending on this decision, the system produces a different type of explanation.
Based on this premise, our Who to Follow and Why framework works as follows: A topical link is usually recommended to person A when person B is authoritative in a topic in which A has demonstrated interest. In this case the explanation takes the form of keywords (e.g., tags on Flickr or hashtags on Twitter) in which A shows interest and B is authoritative.
A social link is recommended when A and B are already part of the same social community, i.e., they have many common contacts. In this instance, the explanation is the standard for current user-recommender systems: A should follow B because they have common acquaintances.
As an important by-product, Who to Follow and Why also implicitly detects communities of users and whether they are social or topical. “Community detection” is arguably the most important topic of study in the analysis of social networks. There are obvious associated sociological interests, but there are also many concrete applications, for example in advertising. Furthermore, our model can describe the detected communities with the most relevant keywords, and for topical ones, can also identify the most influential users.
Though our research is still in its early stages, our framework for the recommendation of topical connections and their correlating explanations could help users discover other users that produce more interesting content than they would have otherwise come across; that type of discovery is much more difficult via current social links. And of course, the benefits for the social platforms themselves could be huge.