KMeans on the Global YouTube Statistics 2023 dataset produces five distinct creator archetypes: mega-scale, mainstream, low-engagement, music-video, and upload machines.

Olist's 2016-2018 marketplace data shows 97 percent single-purchase customers. Revenue growth is all acquisition; retention is near-zero by construction.

Random 5-fold CV makes LightGBM look like the winner on UCI hourly bike-sharing data. Time-series CV says a naive seasonal-mean baseline beats it.

Distance alone predicts 64 percent of trip duration. Adding geography and time of day gets you to 80 percent. The heatmap shows where the missing signal lives.

A stacked Ridge + XGBoost + LightGBM ensemble on Ames houses reaches 0.126 RMSLE, beating a Ridge baseline by about 8 percent. The stack weights are unbalanced, and that's fine.

TF-IDF over Netflix title descriptions and genre tags produces top-10 lists that read like a curator's suggestions. The failure mode is not what I expected.

Cosine similarity on standardised Spotify audio features is the textbook first-pass recommender. Across 2,000 random queries, it lands on the same genre 14.3 percent of the time.

A gradient-boosted price model on the 2019 NYC Airbnb snapshot reaches R^2 0.62 and MAE under $50. Sixty-two percent of the splits are about room type, not location.

The IBM Telco churn dataset has 19 features and one of them does most of the work. SHAP ranks contract type first with a mean absolute impact of 0.88, and the raw churn rate across contract types agrees.

A fraud classifier reaches 0.98 ROC-AUC on the famous ULB dataset and still ships a policy most fraud teams would reject. The 0.5 decision threshold is the bug.