Spot-the-Hits!

PREDICTING THE NEXT VIRAL SONG BASED ON AUDIO ATTRIBUTES: INSIGHTS FROM THE TOP SONGS ON SPOTIFY IN 2024

My capstone project for the Institute of Data aims to predict the popularity of songs on Spotify using a machine learning model trained on a comprehensive dataset of top songs from 70 countries. The dataset includes both audio features and artist popularity metrics, making it a robust resource for building a predictive model.

There were three components required for the capstone encompassing a Jupyter notebook containing the code, technical report, and presentation delivered to my cohort and a panel of expert data trainers.

Components

1. capstoneproject_Mariel Yap_notebook
2. capstoneproject_Mariel Yap_technical_document
3. capstoneproject_Mariel Yap_slides - will need to be downloaded raw

Features Used

The model leverages two key sets of features:

1. Audio Features: Includes attributes such as danceability, energy, loudness, and others extracted from the audio tracks.
2. Artist Popularity: Considers artist popularity metrics like cumulative artist popularity to further enhance the model's accuracy.

Project Structure

• Data Preprocessing: Feature engineering, missing data handling, scaling, and preparation of two datasets.
• Modelling: Various machine learning models were applied, such as CatBoost, RandomForest, LGBM, Gradient Boosting, Lasso, and Linear Regression.
• Evaluation: The models were evaluated on metrics like R² and Mean Absolute Error.

Machine Learning Approach

Several machine learning models were employed, including:

• CatBoostRegressor
• LGBMRegressor
• RandomForestRegressor
• GradientBoostingRegressor
• Lasso Regression
• Linear Regression The final model achieved an R² score of 0.87 when using both audio features and artist popularity, demonstrating its strong predictive performance.

Key Findings

• Audio features alone yielded moderate predictive power, with an R² score of 0.20.
• Incorporating artist popularity significantly improved model performance, achieving an R² score of 0.87, highlighting the importance of artist recognition in predicting song success.

Conclusions

This project successfully built a predictive model for Spotify song popularity, providing valuable insights into the impact of both audio features and artist popularity on a song's likelihood of success.

Future Work

Future improvements could include:

• Expanding the model to include data outside of the Top 50 charts to capture the data of less popular songs.
• Expanding the model to incorporate more detailed artist data, such as social media presence or historical performance trends.
• Experimenting with advanced hyperparameter tuning and stacking models for further optimisation.