Music Recommendation using Spotify Data

Playlists are how most people organize music, and the most useful music recommendation is often the simplest: given the songs already in a playlist, what belongs next? Spotify turned this into the RecSys Challenge 2018 by releasing the Million Playlist Dataset — 1,000,000 user playlists with 2M+ unique tracks from ~300K artists — and asking systems to continue partial playlists under 10 scenarios, from a bare title to 100 seed tracks. The full corpus is too heavy for casual student experimentation, and most tutorials retreat to toy collaborative filtering that ignores the playlist-continuation framing entirely. This project delivers a faithful, runnable slice of the real task: the authentic MPD schema, track embeddings learned from playlist co-occurrence, a three-signal scoring blend, and the challenge's exact evaluation protocol (R-precision, NDCG, clicks) on a documented subset — plus a playlist-builder demo where every recommendation explains itself.

1. MPD JSON slices are parsed into playlist-track pairs; a documented subset is sampled with a fixed seed so any examiner can reproduce the build on a student PC.
2. Track embeddings are learned with a Word2Vec-style objective over the playlist "sentences", capturing which tracks belong together.
3. Audio features (danceability, energy, valence) are normalized per track from the dataset's track metadata.
4. At inference, each candidate track is scored by the three-signal blend against the seed tracks and the playlist title, then ranked.
5. The demo renders the top continuations with score bars and reason chips; the track explorer shows any track's audio-feature profile.
6. Evaluation truncates each test playlist per the 10 challenge scenarios, recommends 500 tracks, and computes R-precision, NDCG and clicks exactly as the challenge defines them.

Tech stack:

• Python 3.10, gensim (Word2Vec-style track embeddings)
• pandas, NumPy, scikit-learn (data pipeline, audio-feature similarity)
• Matplotlib (scenario and metric charts)
• Flask demo app (playlist builder, track explorer, evaluation views)
• Spotify Million Playlist Dataset subset (documented sampling; full 1M optional)

• [Playlist-continuation framing] The task is modeled exactly as RecSys Challenge 2018 defined it: extend a partial playlist with up to 500 ranked tracks, not generic "similar songs".
• [Track embeddings from co-occurrence] A Word2Vec-style objective over playlist-track pairs learns embeddings where tracks that share playlists sit close together.
• [Three-signal scoring blend] Continuation score = 0.5 x co-occurrence similarity + 0.3 x audio-feature similarity + 0.2 x playlist-title matching, with weights documented and tunable.
• [Audio-feature similarity] Danceability, energy and valence profiles let the model match vibe even for tracks that rarely co-occur.
• [Playlist-title matching] Title keywords (focus, workout, party) contribute a cold-start signal when few or no seed tracks exist.
• [Reason chips per track] Every continued track shows why it was chosen — co-occurrence counts, shared vibe, title fit — in the demo UI.
• [Challenge-faithful evaluation] The 10 RecSys scenarios (title-only, first track, first 5, 5 random, first 10/25/100, last 5, title+1, title+5) with R-precision, NDCG and clicks, all computed by the notebook.

• Complete source code (MPD parsing, embedding training, three-signal scorer, evaluation, Flask demo)
• Jupyter notebook (buyer-run procedure: sample subset, train embeddings, evaluate all 10 scenarios)
• Evaluation outputs: per-scenario R-precision, NDCG and clicks tables from your build
• Project report PDF (background, challenge history, methodology, evaluation, dataset caveats)
• PPT presentation for final review
• Viva Q&A preparation document (playlist continuation, embeddings, the three metrics, MPD limitations)
• Setup guide (environment, dataset access via the challenge resources page, subset sampling, running the demo)

• The MPD is sampled with randomization, manual filtering and fictitious tracks added — it is not representative of real Spotify listening, and the report states this plainly.
• Title-only cold start is weak: a brand-new empty playlist has zero track signal and title NLP alone carries the recommendation.
• The corpus covers US playlists from 2010-2017; newer artists, genres and regional tastes are missing.
• The full 1M-playlist corpus is heavy — the student build trains on a documented subset, and full-corpus training is listed as future scope.
• Reported metrics are offline challenge metrics; they do not measure whether real users would actually keep the continued tracks.