UW CSS 581 Machine Learning Project: Personalized Song Recommendation & Audio Feature Identification
In large-scale commercial music recommendation systems, songs are typically recommended based on insight about the songs that other users are listening to who also listened to the songs you’ve listened to. More recently, machine learning algorithms are now being developed with the sophistication of comparing audio features within a song to classify/cluster similar songs. Although these methods are effective in helping determine songs that may be enjoyable to a given listener, song recommenders typically fall short in identifying specifically why a listener may enjoy a given song, and instead rely on similarity detection between songs themselves. Recommendation systems which rely on similar user behavior fall short in that they rarely recommend songs that are less popular and difficult to discover, defeating a core purpose of song recommendation systems.
To solve the problem described above, I developed an MLP Neural Network classification model trained on labeled song audio feature data as a starting point for improved song recommendation tailoring to a given user. As the target user, the binary prediction labels will represent whether I personally like or dislike songs represented by the song audio feature data.