We were seeing a drop in Apple Music session length over three weeks. I ran a cohort analysis segmented by listening context — commute, workout, sleep — and built a gradient boosting model to identify which recommendation features were degrading. I found that the 'New Music Mix' was overweighting recency signals, pushing unfamiliar tracks at high-friction moments. I put the findings in a deck and walked the PM through the model outputs. They ended up deprioritizing the recency weighting in the next sprint.
Apple Music session length was dropping for workout listeners specifically — about fourteen percent over six weeks. The analysis showed 'New Music Mix' was surfacing unfamiliar tracks at exactly the moments users needed energy maintenance, not discovery. Instead of presenting the model, I reframed it as a user moment: 'We are interrupting your run.' I built one slide — a listening timeline showing the skip spike at mile two. The designer immediately recognized the pattern from her own runs. The PM changed the context-aware recommendation logic within the quarter, and workout session length recovered to baseline.