Customize a playlist based on user's feedback on a subset of per-selected songs.
Imagine different songs are different points in a space, our goal is to determine which "segment of space" consists of songs that users like.
Q: How to turn a song into a point?
A: Calculate a song's spectrogram (an image) first. Then we use CNN + autoencoder; the intermediate output of the network, a vector, is essentially a point in an high-dimensional space.
Q: How to segment the area of songs user like?
A: We adopt a technique that is used in graph segmentation tasks. This technique incorporates experts's (your!) feedback to refine the boundary of the space, and this is where 1-bit feedback (like or dislike) comes into play.
We use Bigroom, Blues, Classical, Country, Disco, Dubstep, Hiphop, Lofi, Metal, Techno; all of them are our favorite genres.
The recommendation accuracy is above 70% on Bigroom and Dubstep, both of them are electronic dance music sub-genres.
