Song genre predictor based on Spotify data

The goal of this project is to create a classifier and see how accurately it can predict song genres. Taking a dataset from Spotify [Pandya, 2022], which is al- ready using machine learning algorithms for these purposes, can help assess if the resulting model can be considered apt for a large-scale business or is more appropriate for a smaller audio streaming market player.

Used Machine Learning Methods

• SVM
• Decision Tree
• Gaussian Naive Bayes
• K-nn
• MLP
• Multinomial Naive Bayes
• Nearest Centroids
• Random Forest
• XGBoost

Structure

The contents of the repository are the following:

Folders

• data/ → datasets used for this project
  • spotify_data: the original Spotify Tracks Dataset
  • spotify_clean: dataset without only one genre assigned to each song (generated by using the data-cleaning notebook)
  • spotify_simplified: dataset with only 18 unique genres in total (generated by using the clustering notebook)
  • data_report: exploratory data analysis for the original dataset
• figures/ → figures generated for the presentation and report (generated using the plots notebook)
• ml_methods/ → notebooks with different machine learning algorithms explored for the project

Notebooks

• baseline → implement the majority and rule-based baselines
• clustering → reduce the number of genres in the dataset to only 18 via a combination of agglomerative clustering and manual input
• data-cleaning → choose only one genre for every song in the dataset that appeared with multiple genres
• data-exploration → visualize the features of the dataset and propose preprocessing steps
• hyperparemter-optimization → hyperparameter optimization implemented using GridSearchCV
• plots → generate plots for the report and presentation

Setup

1. Activate your virtual environment
2. Run the following command to install all the dependencies needed for this project:

pip install -r requirements.txt

3. Inspect the code for the different algorithms that were explored (stored under ml_methods/)

Submission details

Team 1

• Elizaveta Nosova (1983805)
• Miguel Samaniego (1980439)
• Nico Sharei (1986818)
• Julian Ament (1981511)
• Artem Bisliouk (1978986)
• Jannik Kranz (1981766)