Instant Search for Music and Podcasts
Spotify boasts of 70 million songs and 3.2 million podcasts on its platform. When we click on that search button, how does it ensure we get what we expect? We would be overwhelmed by the options if it were just a textual match. In this blog post
, Spotify explains how it defines its search algorithm.
There are three essential requirements of a good search algorithm. First, it is instantaneous, i.e. results update with each keystroke. Second, it is heterogeneous, i.e. the user could be searching for music or podcast. Third, it understands multiple representations of the same word, i.e. text in context.
In general, Spotify found that users delete 53% more characters when searching for a podcast than for a song, even though the average query length is the same. Furthermore, upon searching, users download podcasts nearly six times as many songs. These insights led them to develop a neural architecture called Neural Instant Search (NIS)
, that outperformed existing algorithms.