Hit Song Prediction Through Machine Learning and Spotify Data

This study predicts hit songs using metadata from the Spotify API[8]. The dataset includes over 20 genres, each with 40 songs, equally divided between hits and flops, gathered using spotipy[7]. Prediction is based on the popularity feature, rated from 0-100. Models were trained on features like danceability, energy, loud-ness, speechiness, valence, and tempo. The dataset was split using train_test_split (10%, 20%, 33%) and kfold cross-validation with k val-ues of 2, 5, and 10. Models were trained, evaluated, and tested, with kfold cross-validation showing the best accuracy and the least over-fitting. Scikit-learn’s classifiers, ensemble models, and MLPClas-sifier were used, with PassiveAggressiveClassifier and AdaBoost showing 60% accuracy. Ensemble methods like extra trees and ran-dom forest, along with neural networks, performed well. Gaussian Process, Naive Bayes, and ridge classifiers stood out among stan-dard models. These results suggest that enhanced models, especially neural networks and decision tree ensembles, could improve hit prediction. Future work may explore frequency and lyric analysis.