Can we use the words in song lyrics to accurately predict a song's genre?
- Languages & Libraries: Python, Pandas, NumPy, Scikit-Learn, XGBoost, Seaborn, Matplotlib
- Tools: Jupyter Notebook, Git, SQLite
Using the Million Song Dataset and several corollaries, I uses ground truth genre classes to build machine learning models to assist with genre classification.
This project is a deep dive into the intersection of Natural Language Processing and music. I developed a machine learning pipeline to classify songs into ten distinct genres based solely on the frequency of words in their lyrics. The result is a robust model that demonstrates the power of lyrical content in defining musical style.
A selection of word clouds showing the most distinctive words for various genres.
This capstone project successfully demonstrates an end-to-end data science workflow, from data ingestion and cleaning to feature engineering, modeling, and evaluation. The best-performing model, an XGBoost Classifier, achieved a promising level of accuracy, proving the viability of using lyric word counts for genre prediction. The HistGradientBoostingClassifier fared slightly worse on these metrics but had a much better confusion matrix which showed more promise for a future working overall genre classfier.
A slide deck overview of the project is available here: Lyrical Genre Classification Presentation
This project followed a structured data science methodology to ensure robust and reproducible results.
The foundation of this analysis is a combination of three powerful datasets:
- The Million Song Dataset (MSD): Provided the core track metadata.
- The musiXmatch Dataset: The official lyrics collection for the MSD, containing a bag-of-words representation for over 200,000 tracks.
- The Top-MAGD Dataset: Used to map tracks to their top-level musical genres.
These datasets were meticulously cleaned, merged, and processed to create a unified dataset ready for analysis.
Before modeling, I conducted a thorough EDA to understand the data's characteristics and uncover initial insights. This included:
- Genre Distribution Analysis: Visualizing the frequency of each of the 10 genres in the dataset to understand class balance.
- Distinctive Word Analysis: Generating word clouds for each genre to visually identify the lyrical features that make each style unique. This step was crucial for confirming that lyrical content held a strong enough signal for classification.
The core of the project was to build and compare multiple classification models to find the most effective approach.
-
Feature Set: The primary features were the top 5,000 most frequent words from the musiXmatch dataset, forming a classic Bag-of-Words model. Additional features like song duration and release year were also included.
-
Models Tested: I implemented and evaluated a suite of powerful classification algorithms:
- Logistic Regression
- Decision Tree
- Random Forest
- K-Nearest Neighbors
- Gaussian Naive-Bayes
- HistGradientBoostingClassifier
- XGBoost Classifier
-
Performance: The models were evaluated on their accuracy and F1-score. The tree-based ensemble methods demonstrated the strongest performance.
| Model | Test Accuracy | Test F1-Score |
|---|---|---|
| XGBoost | 42.5% | 40.3% |
| HistGradientBoostingClassifier | 41.0% | 40.0% |
| Logistic Regression | 37.0% | 35.5% |
| Random Forest | 38.3% | 32.2% |
The confusion matrix for the XGBoost model.
The confusion matrix for the HistGradientBoostingClassifier model.
This project is licensed under the MIT License - see the LICENSE.md file for details.
- Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian Whitman, and Paul Lamere. The Million Song Dataset. In Proceedings of the 12th International Society for Music Information Retrieval Conference (ISMIR 2011), 2011.