A Robust Approach to Predict the Popularity of Songs by Identifying Appropriate Properties

Abstract

Hit Song Science is a major research topic which is being discussed today in the eld of Music Information Retrieval. In order to identify and predict whether a song could be a hit song or not is yet a challenging task. This thesis investigates the ability of using machine learning to make predictions on Sinhala songs whether they would be a hit or not. More than 13,000 Sinhala songs were collected by web scraping in a popular Sinhala music website which also contributes by presenting a dataset that can be used for further research purposes. The number of downloads and the view counts were used to derive a equation to measure the popularity. The features extracted of each song is used by the XGBoost classi cation algorithm. The songs are initially grouped into 3 classes based on their popularity, and later by performing machine learning algorithms on a set of features extracted on each song, the impact of the Linear Predictive Coding (LPC) overall average (LOA) and Mel-Frequency Cepstral Coe cients (MFCC) features, to the end result is indicated by the use of SHAP process.