
While we’ve developed advanced machine-learning systems like Pinnability that predict how relevant an idea is to a Pinner, the first challenge is figuring out which of the more than 100 billion Pins to even consider since we can’t possibly score them all at once. This is why having a graph-based recommendation system is valuable to us. Pixie solves the candidate generation problem by starting graph traversal from a set of nodes we already know are currently relevant to the Pinner. Then, it only examines the portion of the graph nearest to these nodes by using a biased random walk…
…r things you are capable of paying attention to than just what makes the most noise on the surface. Just because a quiet room doesn’t scream with excitement like the idea of immersing yourself in a movie or a TV show doesn’t mean that there isn’t depth to explore there.