Will you actually like this track

Level 3: a month or more, real users. Use with: clean-a-dataset, train-a-model, simple-dashboard. Resume line when done: Built {project_name}, a music-taste model trained on my full listening history and scored on future listening — at {url}

Brief

Request your full listening-history export and build a model that predicts whether you will like a track you have not heard yet. The hard parts are honest ones: deciding what "liked" even means in your data, keeping features that leak the answer out of the model, and testing on the future instead of the past. The demo is a dashboard that scores incoming tracks and then keeps a public tally of how the predictions held up against your actual listening.

Personalize

• Your export. Your own streaming history export, the full one (most services include play counts, skips, and timestamps). While you wait for it, write down your guess about what predicts your taste.
• Your label. Define "liked" from behavior, and defend it: saved, or replayed more than N times in the following month, vs skipped early and never returned. The write-up must discuss what this definition gets wrong.
• Your baselines. Majority class, then a one-liner: "liked if the artist is already in my top 50." On personal listening the artist rule is brutal to beat, which is exactly why it goes on the scoreboard.

Milestones

1. Get the export, load it, and profile it: span, play counts, skip data, gaps. Chart your listening over time before modeling.
2. Write the label definition and apply it. Report the class balance, and revisit the definition if 95 percent of tracks land on one side.
3. Split by time: the most recent months are the test set, frozen. Explain in the README why a random split would let the model see the future.
4. Build features from track and artist metadata (an API helps here). Audit each feature for leakage: anything computed from listens after the track first appeared is cheating. Document what you excluded and why.
5. Score both baselines on the test set.
6. Train a simple classifier on the past-only training data.
7. Evaluate on the frozen test set: accuracy and precision on the "liked" class, next to both baselines.
8. Read the worst errors: tracks the model was sure about and got wrong. Write what the features could not capture (mood, a friend's recommendation, one great chorus).
9. Build the dashboard: paste or pick a track, get a like probability, with the evaluation table on the page.
10. Run the prospective test: score 20 tracks you have not heard, listen over two weeks, log real outcomes next to predictions.
11. Publish the final tally and the limitations: your label is a proxy, taste drifts, and two weeks is a small window.

Done means

• Model trained only on data from before the test period, with the temporal split explained in your words
• Leakage audit written down: which features were excluded and why
• Test accuracy and liked-class precision published next to the majority and top-artist baselines
• Prospective test completed: at least 20 predictions logged before listening, real outcomes published next to them
• Dashboard live with the scoreboard and the worst misses shown
• Label definition and its flaws discussed openly; no claim that the model knows your taste, only what the tally shows