WTF: The Who to Follow Service at Twitter
notes based on the paper https://stanford.edu/~rezab/papers/wtf_overview.pdf
1. Personalized PageRank
This algorithm starts from a specific user and spreads out like water flowing through a network. It assigns higher importance to nearby users (closer connections) while occasionally jumping to random users in the network.
Example: Imagine you're User A. Personalized PageRank begins at your profile and randomly visits your connections—friends, followed accounts, and so on. If you follow User B and User C, and they both follow User D, the algorithm will identify User D as someone worth recommending to you because they’re indirectly connected through multiple paths.
2. SALSA (Stochastic Approach for Link-Structure Analysis)
This algorithm creates two groups in the graph: "hubs" and "authorities." It alternates between the two, hopping back and forth to find users worth recommending. A hub is like a directory pointing to interesting people (authorities), and authorities are the interesting people themselves.
Example: You (User A) follow 10 accounts related to technology. Those accounts also follow big tech influencers. The algorithm connects these two sides—the accounts you follow ("hubs") and their followed accounts ("authorities")—and recommends these influential tech profiles to you.
3. Most Common Neighbors
This algorithm suggests people who are popular among your direct connections. If a lot of your friends are following someone, the algorithm assumes you might want to follow that person too.
Example: If User A follows User B, C, and D, and all three of them follow User E, then User E will likely be recommended to you. It works on the idea that if many of your friends find someone interesting, you might too.
4. Closure Algorithm
This method looks for patterns where connections can be completed to form a loop or triad. It tries to close open connections in your network by suggesting users to form relationships.
Example: Let’s say User A follows User B and User C, and User C already follows User A. The algorithm will suggest User B to User C, completing the triangle. It’s like saying, “You three are in the same circle—why not follow each other?”
5. Sim(Followings)
This algorithm looks at the people you follow and finds others who are similar to them. It assumes that if you follow User B, and they follow similar people as User C, then User C is worth recommending.
Example: If you follow User X, who is into fitness, and User Y is also followed by most fitness enthusiasts in your network, the algorithm suggests User Y to you because they share a lot in common with User X.
6. Circle of Trust
This method builds a "personal bubble" of trusted connections around you using random walks. It repeatedly starts from you and hops across your connections to create a personalized view of the network.
Example: Imagine User A follows 20 people. The algorithm “walks” through these connections, exploring their connections and beyond. If many walks end up frequently on User D’s profile, it concludes User D is significant for you and recommends them.
