03 · Machine learning / Case study
Spotify popularity and genre modelling
A reproducible regression and classification pipeline with baselines, feature engineering, validation and automated tests.
The problem
Separate predictive signal from an appealing but weakly generalising story about music data.
My contribution
I built the music-data pipeline: year cleaning, a decade feature, baseline comparisons, model selection, metric reporting, diagnostic figures and tests. Popularity regression and genre classification have separate models and evaluation criteria.
Two questions need two measures
Popularity is a numeric target evaluated with RMSE. Genre classification has imbalanced classes, so a single accuracy value needs context.
Benchmark and validation are different datasets
CatBoost: benchmark RMSE 6.52657; quick local validation RMSE 9.8633. Linear SVC: benchmark accuracy 0.57142; local validation 0.5875. These are not results from one evaluation split.
More complexity does not always help
The TF-IDF text variant remained an experiment. The default tabular Linear SVC gave the stronger reproducible validation result and matched the validated prediction file.
Evidence and interpretation
What the result actually says.
Reported benchmark RMSE of 6.52657 and classification accuracy of 0.57142.
The project is valuable for the model-selection decisions it documents. A benchmark score does not replace validation, and extra features must earn their place.
Results depend on the dataset, split and target scale. The work documents overfitting risk and is not a production recommender.
The next useful experiment
Broaden validation, inspect rare genres and address high-cardinality categorical risks. This pipeline is not a music recommender or a forecast of commercial success.
Descriptions checked against the public project documentation on 17 September 2026. Scores are reported project results; the experiments were not rerun for this website update.