Introduced by Vivek Jain, Katharine Jiang, Natalia Layson, Rishabh Parikh and Regina Xu
Songs have many different attributes attached to them, which make it hard to classify what songs a person might listen to based on their mood at the time. Although happy and sad emotions are the same for most people, the songs they listen to vary across the spectrum.
In this project, a team of UC Berkeley students aimed to classify songs as happy or sad through a variety of different measurements, including valence, energy, beats per minute, and the lyrics of the song itself. With various classification models, they determined what type of songs people listen to in different moods using a dataset that combines features from Spotify and lyrics in a bag-of-words format.
The results gave insights on the top 10 words that occurred for each mood, as well as feature distributions and prediction accuracies for different classification models. The best classifier was Random Forest on the dataset that used all of the features provided by Spotify. The general public can use this in a potential interface featuring a “weekly mood report” Spotify or another music app.