Amelie Anglade will speak at The Realtime Conference Europe
We're thrilled to resume our speaker's announcements with Amélie.
[photo courtesy of Thomas Bonte]
Amélie, who is currently at SoundCloud will tell us about DiscoRank: a smart algorithm used to sort users, tracks, sets... but she describes it better than I do!
When Alexander Ljung presented the Next version of SoundCloud in December 2012, he introduced our improved search and the brand new Explore section. Those two features share a common algorithm to boost relevant results: the DiscoRank. Inspired by the PageRank, it extends its principle to the full graph of searchable entities (users, tracks, sets, etc.) in which each activity on SoundCloud is an edge connecting two nodes. The discoverability ranking of each node is then computed using the weighted graph of entities and activities. > We are dealing with millions of nodes and edges so in our first implementations the DiscoRank was taking several hours to converge. We trimmed the computation down to a few minutes. To that end we optimized how the full graph is stored in memory, with a tailored sparse representation. Using versioned copies of the DiscoRank we can both swap between them when testing things out and compute the most recent DiscoRank starting from results of previous runs, only adding the new activities to the graph.
In this talk we will describe the DiscoRank algorithm and how it is integrated to search backend and adapted to the time-sensitive requirements of the Explore section. In particular we will elaborate on how it was optimized for speed and memory size to serve the user sounds that emerge from the latest trends.
