Classification of Music for Study Based on Spotify Audio Features Using Random Forest with Feature Importance Analysis and Reduction

Abstract

Music has a significant impact on the way a person thinks and feels in their daily activities. This study aims to categorize the types of music that are suitable for learning activities by using Spotify's audio feature, to create a more flexible and personalized music recommendation system. The dataset used comes from Spotify Study Music which consists of 172,819 songs with 12 audio features, which are grouped into three main categories, namely Pop tracks, Classical soundtracks, and Lo-fi tracks. The research process includes data pre-processing, handling class imbalances using SMOTE, data normalization, feature significance Analysis, Cross Validation, and feature reduction. Normalization results show that all features have been in the range of 0.0-1.0 without changing the characteristics of the original distribution. The Random Forest Model performed exceptionally well with an average accuracy rate of 99% on cross-validation and 99.9% on training data, indicating the model's ability to efficiently recognize musical patterns. Important Feature Analysis shows that energy, loudness, acousticness, instrumentalness, and liveness have the most significant influence in distinguishing music characteristics for learning, while mode, popularity, duration_ms, and danceability when removed using Feature Reduction analysis show a significant decrease in accuracy. This study recommends maintaining the features of acousticness, instrumentalness, and liveness because it plays an important role in maintaining the stability and accuracy of music classification models that support the learning process.