Our work has demonstrated that matrix factorization is an effective way to create a recommender system--not just for movies or shopping items, but also for song suggestions for user generated playlists. Even with our too sparse matrix, we managed to get recommendations that fit well with our playlists, even across many different genres.
Given more time, a better way to quantify our results would have been to conduct PCA on audio features taken from Spotify's API for the songs that we recommend and the songs in the playlist to see if the recommended songs cluster together with the rest of the playlist. However, the Spotify API is difficult to query for many songs, and due to storage and RAM limitations, we could not extract all of the audio features and perform a quantitative analysis on the performance of our model other than by precision@K.
Though we got great results with just the preferences of the users in the playlists, it would have also been interesting to look at other features of the songs, including the audio features mentioned above, but also lyrical content to provide information about another way that people may construct playlists. This may be another way to suggest songs in playlists that aren't limited to just one genre, which our model made reasonable suggestions for, but mostly from the set of most popular likely choices.