Back to research
CSE422: Artificial Intelligence · 2023

Song Virality Prediction Using ML

An end-to-end pipeline over 114,002 Spotify tracks that defines 'viral' by clustering and compares five classifiers, and a useful lesson in target leakage.

View on GitHub
[ ARC ]

Architecture

  1. 01

    Preparation

    • Median / mode imputation
    • IQR outlier capping
    • One-hot genres, scaling

  2. 02

    Label by clustering

    • K-Means (k = 3)
    • Popularity, danceability, energy
    • Top cluster = viral

  3. 03

    Classifiers

    • KNN, decision tree, LogReg
    • Random forest, naive Bayes

  4. 04

    Evaluation

    • Accuracy, F1, ROC-AUC
[ KEY ]

Technical highlights

  1. 01

    Built the full pipeline: cleaning, feature engineering, clustering-based labelling, and a five-model comparison.

[ LIM ]

Limits & lessons

  • The models scored near 100%, but the label was defined from popularity, danceability, and energy, which were also inputs. That near-perfect score measures target leakage, not virality. The next version should define virality from signals the model can't see.
[ STK ]

Stack

  • Python
  • scikit-learn
  • K-Means
  • pandas
  • Matplotlib