Analyzing Spotify Song Features Using K-Means Clustering and PCA Visualization

This project explores how to intelligently group Spotify songs using machine learning techniques. It primarily uses the K-Means clustering algorithm to analyze audio features and achieves data visualization through PCA dimensionality reduction. It aims to solve the problem of rough genre labeling by traditional manual methods and provide data-driven refined solutions for music recommendation systems and personalized services. The original author of the project is Luis7ml, published on GitHub (link: https://github.com/Luis7ml/Spotify-Songs-Clustering-with-K-Means-and-PCA) on June 2, 2026.

In the era of music streaming, platforms like Spotify process massive amounts of song data daily. Understanding the intrinsic features of songs and discovering connections between similar music are key challenges for recommendation systems. Traditional manual genre labeling is rough and cannot capture subtle audio differences. Machine learning techniques can automatically discover song similarities, provide more refined classifications, and complement subjective classification systems.

K-Means Clustering Algorithm

As an unsupervised learning algorithm, its core is to divide the dataset into K clusters, making the similarity within the same cluster high and the difference between clusters large. In the project:

1. Extract quantitative audio features such as rhythm intensity, pitch variation, energy, etc.
2. Calculate the Euclidean distance of feature vectors to measure similarity.
3. Iteratively optimize cluster centers until convergence.

PCA Dimensionality Reduction

Since audio features are multi-dimensional (10+ dimensions), PCA maps high-dimensional data to a low-dimensional space through linear transformation, preserves variance information, and enables 2D/3D visualization for easy observation of cluster distribution.

1. Data Preprocessing: Standardize audio features of different ranges to avoid numerical dominance in clustering;
2. K Value Selection: Use the elbow method and silhouette coefficient to evaluate different K values and select the optimal number of clusters;
3. Feature Engineering: Combine/transform original features (e.g., energy + danceability to identify music suitable for dancing);
4. Result Interpretation: Analyze the mean values of features for each cluster and attach interpretable labels to the clustered groups.

The practical value of this clustering method includes:

• Personalized Recommendation: Identify user preference clusters and recommend songs from the same cluster to improve accuracy;
• Automatic Playlists: Create themed playlists based on clustering (e.g., fast-paced sports music, relaxing soft music);
• Music Discovery: Help users find similar but unheard new music;
• Industry Analysis: Record companies/musicians analyze popular trend features to guide creation.

Possible expansion directions for the project:

• Combine deep learning (e.g., autoencoders) for nonlinear dimensionality reduction to capture complex patterns;
• Introduce time series analysis to study the evolution of music popularity trends;
• Integrate lyric text analysis to achieve comprehensive understanding by combining audio and semantics;
• Develop a real-time clustering API to support streaming online recommendations.

This project demonstrates the application potential of machine learning in the music field. Through the combination of K-Means and PCA, it automatically discovers song similarities and intuitively understands the structure of music data. Its value lies in providing an objective data-driven way to understand music and complementing subjective classifications. For enthusiasts, it is a tool to discover new music; for developers, it is the foundation for intelligent recommendations; for researchers, it is a new perspective to analyze trends. With technological progress in the future, there will be more accurate and personalized music services.