Harvard CS109A, fall 2018
Final Prject Group No. 32
Alkmini Chalofti, Liam Corrigan, Andra Fehmiu, Shih-Yi Tseng
Today there is a prevalence of modern music streaming applications like Spotify, Pandora, and Apple Music, where millions of users consistently are consistently generating playlists. One aspect of retaining customers is recommending new songs so that the user will be engaged in discovering new artists. The aim of our project is to create a playlist recommendation system, whereby given some set of songs in a playlist we would generate some distinct set of songs that a user would be likely to enjoy given the songs on that original playlist.
In creating our playlist recommender we had access to three main datasets, the Spotify Million Playlist dataset, the last.fm Million Song Dataset, and the Spotify API which provides quantitative metrics about a given track such as tempo. In this website we describe the creation of our model starting with our Exploratory Data Analysis (EDA), and the going over a review of the literature on recommender systems. We then demonstrate the Singular Value Decomposition (SVD), k-means clustering, and Neural Network Collaborative Filtering (NNCF) models that we used, and discuss their results. Finally, we present an example of playlist generation in action, to provide an intuitive sense for what our model does.