Spotify Track Popularity Prediction with Statistical Learning

This project analyzes Spotify track-level audio features and genre information to study what drives track popularity and to build predictive models for popularity scores.

The work combines exploratory data analysis, statistical testing, feature-based modeling, and machine learning methods to understand the relationship between track characteristics and popularity.

Start here: Final Report PDF

Project Overview

The analysis focuses on a Spotify dataset containing audio and metadata features such as:

• valence
• acousticness
• danceability
• explicit flag
• track genre
• loudness
• popularity

The project uses these features to:

1. explore trends and relationships in the dataset
2. test statistical patterns across genres and audio characteristics
3. build predictive models for popularity
4. compare regression-based and support vector methods
5. evaluate model performance using standard regression metrics

Dataset Scope

• Original source material covered approximately 90,000 unique tracks across more than 100 genres
• The final statistical analysis focused on 10 popular genres for clearer comparison and modeling
• Popularity was modeled as the target variable on a 0 to 100 scale

Methods and Tools

• Language: Python
• Environment: Jupyter Notebook
• Libraries: pandas, numpy, matplotlib, seaborn, scipy, statsmodels, scikit-learn, tabulate
• Modeling approaches: linear regression, support vector regression, hyperparameter tuning with GridSearchCV
• Evaluation metrics: R-squared, mean squared error

Statistical Findings

• Popularity differed significantly across genres based on the Kruskal-Wallis test
• Popularity also differed significantly between explicit and non-explicit tracks based on the Mann-Whitney test
• Genre and explicitness showed dependency in the chi-square analysis
• Acousticness showed a strong positive relationship with popularity
• Loudness showed a strong negative relationship with popularity
• Valence had only a weak relationship with popularity

Model Results

Model	R-squared	Mean Squared Error
Multiple Linear Regression	0.768	82.83
Lasso Regression	0.767	83.24
Support Vector Regression	0.742	92.33

The linear and lasso models performed best in the final comparison, while SVR captured similar directional effects but with lower overall predictive accuracy.

Repository Structure

spotify-track-popularity-prediction-statistical-learning/
  README.md
  .gitignore
  requirements.txt
  data/
    README.md
    raw/
      spotify-track-popularity-dataset.csv
  notebooks/
    spotify-track-popularity-prediction.ipynb
  reports/
    final-report.pdf
    final-presentation.pdf
    final-presentation.pptx

Included Files

• notebooks/spotify-track-popularity-prediction.ipynb: final cleaned analysis notebook
• data/raw/spotify-track-popularity-dataset.csv: working dataset used in the project
• reports/final-report.pdf: final written report
• reports/final-presentation.pdf: presentation summary of the project
• reports/final-presentation.pptx: editable final presentation deck

Main Conclusions

• Tracks with higher acousticness and lower loudness were more likely to be popular in this dataset
• Genre and explicitness were statistically important predictors and should not be ignored in modeling
• Multiple Linear Regression and Lasso Regression offered the best balance of interpretability and performance in the final project
• The project shows how statistical testing and machine learning can be combined in a practical music analytics workflow

Notes

• The staged GitHub version keeps the final notebook and strongest supporting material.
• Draft notebooks and intermediate versions are intentionally excluded from the public repo.
• The notebook has been adjusted for repository-relative data access instead of Google Drive-only paths.
• The report and final presentation are the primary polished deliverables for this repository.