Role of Search Algorithms in Spotify’s Discover Weekly
Companies like Pandora Radio have made a name for themselves through their promise of finding and offering new, exciting music to their users. Yet Pandora, which devotes itself entirely to this purpose, has seemingly taken a backseat to the (much) newer Spotify Discover Weekly series, which, according to Nikhil Sonnad’s The Magic that Makes Spotify’s Discover Weekly Playlists so Damn Good, streamed 1.7 billion songs to Spotify Users in the 6 months following its introduction. As a personal fan of these playlists (and a New Yorker – the service is now broadly advertised in subway trains and stations), I was interested in the algorithm behind the personalized playlists.
While Sonnad’s articles do not explicitly mention networks, crawling, or ranking, the underlying ideas behind Spotify’s suggestions are analogous to the ways in which search engines rank web pages. The most important factor weighed into the playlist generation is the playlist a user already listens to. As Sonnad says, “[Spotify] gives extra wright to the company’s own playlists and those with more followers.” Weighing in the number of followers a playlist has gives a playlist credibility, which can be quantified in the same way we assign “hubs” a value that factors into the value associated with their endorsed “authorities.” The algorithm also factors in the “links” between a user’s favorite songs: “if your favorite songs tend to appear on playlists along with a third song you haven’t heard before, [the algorithm] will suggest the new song to you.” This, too, bears a resemblance to the type of indirect endorsement provided by hubs (which, in this example, would be your “favorite songs,” while the corresponding authority would be the suggested songs which appear in the same playlist).
Spotify also consistently avoids suggesting music in genre’s radically different from those that are into, much like the search ranking process described in the “PageRank” video in the Web Search module filtered out Youtube and Amazon when searching for museums. The connection between Google’s search engine and Spotify’s music suggestions potentially manifest themselves in Sander Dieleman, a researcher for Google’s AI branch who formerly interned at Spotify (though, to be fair, Sonnad doesn’t explicitly state Dieleman contributed to Discover Weekly).
Yet, like any search engine, Spotify’s Discover Weekly is not without its shortcomings – sometimes, it can provide songs users despise. Yet even these have value: when users skip a song within thirty seconds, Spotify interprets the act as a negative response to the song, and is then less likely to suggest similar songs. In this way, sending in bad songs as a “test” can help to reinforce the perceived value of other songs and genres. Adjusting to these skips is analogous to the example in the module videos in which if users of a search engine regularly click on the second link of the results provided by some constant query, the second link is more likely to be moved up a position (to occupy the first slot in the search results).
As demonstrated by its implementation in the Discovery Weekly series, the power of network crawling (and ranking) has applications far beyond search engines. As services that rely on humans as their products (like Facebook, Spotify, etc.) continue to form and grow, and as more data becomes available to these companies, efficiently and accurately analyzing networks to find meaningful relations between users and content is becoming increasingly important.