Music consumption as a mental-health signal

Self-reported mental health data is sparse and arrives late. Streaming behaviour, by contrast, is continuous and passively collected, which makes it an interesting candidate signal if it carries any real predictive weight.

Preprocessing ran on Databricks with Spark to handle the volume. I compared gradient-boosted trees, logistic regression and random forests, tracking every run through MLflow so the comparison stayed honest rather than becoming a search for the best-looking number.

The headline result is a modest one, and that is the finding. Listening habits carry signal, but not enough to stand alone. The threshold optimisation mattered more than the model choice, which is the usual story in imbalanced classification and the part most write-ups skip.

Results

• AUC of 0.676 across the model comparison
• F1 raised to 0.747 through threshold optimisation
• Full run tracking in MLflow