Results

See below left for the original playlist, and right for the recommended songs. Analysis of the recommended songs can be seen further at the bottom.

Result Interpretation

With the playlist on the left, our ANN model was able to learn the tags for the playlist and make predictions based on the tags. The regular songs recommended are the songs come from a random generated sample of the Million Song Database. In order to recommend newly released songs and songs that are unpopular, we got another sample from the API. And with the separate sample, we run the same algorithm to get recommended newly released songs and unpopular songs.

A point to notice is that the unpopular song we got from the API is defined by the popularity at the current time. So in the result here, we have "I want it that way" from Backstreet Boys. It only means that fewer people are listening to that song right now.

Looking at the recommendations from New Released Songs, we can see that songs recommended here are relatively new (i.e. "Sundays" by Emotional Oranges was released Oct. 2019) and the simply by listening to the music on both sides, you can see that the acoustic features feel the similar from a user's standpoint. We have asked several users (n = 5) to play the original playlist of songs and the recommended songs as a rudimentary user testing, and they all agreed that acoustically, they both share similar acoustic features and therefore validating the overall quality of recommended songs from the model we built.