Hit Song Prediction: Using Ant Colony Optimization for Feature Selection
Hit Song Science is a binary classification problem which aims at anticipating the success of a song before its release. This is particularly useful to identify talented singers, musicians, lyricists etc. before they get contracts from Music labels. In this work, we considered the Hit Song Science problem as a classification problem and attempted to solve it using a feature selection technique known as Ant colony Optimization to identify most prominent and definitive features and improve upon the accuracy of previous work done on this problem statement. We have used Spotify Web API to extract acoustic features of almost 6000 songs accumulated from online repositories. We test multiple classification models such as XG Boost, Random Forest etc. on our dataset. The results demonstrate that, as compared to existing approaches, our approach selects a minimal number of features and achieves better performance.