Predicting Hit Songs: An Exploratory and Comparative Study

Data science is currently at the forefront of the music industry, especially in streaming media companies like Spotify as songs’ audio features can be labeled with specific values and weights. Using the Kaggle Spotify Tracks Dataset, a labeled dataset in which hits are defined as tracks at or above the 80th-percentile popularity is assembled to train the models. A comprehensive exploratory data analysis characterizes the nonlinear distributions and indicates the correlations between audio features and song popularity. Four classifiers—Logistic Regression, Support Vector Machine, Random Forest, and Neural Network—are trained in standardized pipelines. To align with the discovery objective of optimizing the genuine “hit” prediction, the study applied class-weight balancing and cost-sensitive threshold tuning to maximize recall. The results demonstrate that a tree-based ensemble, particularly Random Forest, gives the best precision-recall balance, while SVM also outputs a competitive recall under a tuned threshold. The study explores the connection between audio features from an objective statistical standpoint and identifies that tree-based ensembles perform better when dealing with tabular datasets.