COLLABORATIVE FILTERING

Thus, we used “cosine” as our metric in our KN_model. Given the sparse matrix, this model would return the n nearest neighboring playlists, from which we could make our song suggestions.
Once we had the n nearest neighboring playlists, we iterated through every song in each neighboring playlist and returned the desired number of non-overlapping songs based on two factors: playlist similarity score and song occurrence frequency. Playlist similarity was based on the notion that our model should favor selecting songs from playlists that were scored as more similar to the original playlist. Thus, the candidate songs were given a score of 1/i, where i is the index of its playlist as ordered by similarity. This method assumes that the difference in similarity from Playlists 4 and 5 vs. Playlists 1 and 2 is proportional to the difference between ¼ and ⅕ vs. 1 and ½, which may not be the case, but this inverse method was a suitable metric for us to score and rank the candidate songs.
The other factor we used was song occurrence frequency, which ensured that songs that appeared in multiple similar playlists would be considered as better recommendations. To implement this, we added the playlist similarity scores for each occurrence of a song to calculate the final score for each song. Then, the desired number of song recommendations in order of highest score was returned to the user. The model was trained using a 30-70 test-train split on 100 randomly selected playlists from the million playlists dataset. This choice of sample size was mainly due to the limitations in the computing power/time of our laptops.