Analyzing the Friend Graph
A few days ago, on September 18th, 2014, Facebook announced that they have made a few tweaks to their new feed algorithm. One of the key parts of keeping Facebook relevant in the pop culture is to keep the content that it presents to a user relevant. As you probably know, Facebook tries to put relevant posts from friends or pages that the user is associated with. This of course raises the question, what is relevant? How is relevance determined? Two of the most important things are closeness and amount of interaction, that is, how closely connected a user is to a person or page posting something, and how many people interact with that post through things like “liking” the post or commenting on it. Simply put, if many friends of a user interact with a post, Facebook will think it is relevant and important, and so it will have a higher priority when being presented to the user. The problem Facebook is addressing though, is that sometimes the post itself is time-specific and becomes irrelevant very quickly, but the interactions make Facebook think that it is still relevant. So on the surface it seems to make sense, lots of friends commenting or liking a post should mean it is something the user would want to see, but this can backfire.
Facebook’s new algorithm is aimed to resolve this problem by looking at trending posts. That is, when Facebook decides what is relevant to the user, it will also take into account whether others are posting about the same topic on Facebook, even people completely unrelated to the user. Topics that are being mentioned in posts, or trending topics, are more likely to show up while other users keep writing new posts about that topic. This takes the good old fashioned “friend graph,” and adds some new factors. It still looks at what posts are popular among friends, but when it weighs the relevance of the post to a given user, it now also applies data collected from the entire Facebook community as opposed to the small subsection of the graph that is a user’s friends.
The algorithm change also looks at the timing of the post. If interactions with the post quickly drop, then it is assumed to only be relevant for a short period of time. Similar to the first change this adds more complexity to the way that the relevance of new posts are weighed in addition to the original method of analyzing the graph of a user’s friends to determine what is relevant.
I think its interesting to look at how in certain ways the original idea behind how the relevance of posts are weighted can lead to irrelevant or uninteresting posts appearing on a user’s news feed. It seems quite intuitive, if a friend with a strong tie to the user posts something, generally it is more relevant, or if many of the user’s friends are interacting with the post, intuitively we would think it is interesting to the user. This update brings an interesting insight as to how the friend graph of a user can be analyzed, and how there is so much more complexity that needs to be accounted for other then how close the user is to a given friend. Particularly this idea of taking a step back, looking at the big picture, that being the posts going through Facebook as a whole, and taking that into consideration. For example, if there is strong triadic closure between the user, friend A, and friend B, if friend A posts something, and friend B comments, it should be relevant to the user, particularly if more close friends interact with the post. But lets say the post is about something very time-specific, like some sort of intermediate update on an event. It is probably less interesting after said event relative then say, a post about a new iPhone the day it comes out that is being talked about throughout Facebook. In general, with the ability to analyze data to tailor it to an individual, the idea that one could try too hard, and end up needing to generalize the user just a bit, seems to violate so many ideas about how the friend graph of a user ties into the relevance of friends’ posts. It’s interesting to look at Facebook and see it like a giant real-world experiment where we can really look at various theories related to graphs and sometimes discover that sometimes things can seem rather counterintuitive, like the one that this algorithm tweak accounts for.