Sonic, A Disciplined Music Recommender

Collaborative filtering on Last.fm, evaluated the way a research team would ship it.

This is the written companion to the Sonic recommender. It documents what the system does, how it was built, and. Most importantly, how we know the numbers are real: a frozen evaluation harness, a leakage-safe holdout, strong baselines, and significance tests. The served model is EASE, a linear item-item autoencoder.

The one-paragraph version

We recommend musical artists from implicit listening data (play counts, not ratings). Starting on the small Last.fm-2k research set, a disciplined hyperparameter search found that a low-capacity matrix-factorization model won and a deep variational autoencoder lost, the textbook result for tiny, sparse data. We then scaled to Last.fm-360K (real, uncapped histories, 1.68M interactions), and the ranking flipped: the higher-capacity EASE autoencoder now beats tuned ALS by a statistically significant margin, and it is the served model. Under honest full-catalogue ranking it scores NDCG@10 = 0.22, NDCG@100 = 0.36, Recall@50 = 0.42. Squarely in the published state-of-the-art band for music recommendation.

Headline results

0.219

NDCG@10, tightest cutoff

0.361

NDCG@100, SOTA band

0.423

Recall@50, matches MSD SOTA

+0.036

EASE − ALS, p < 0.001

All metrics are full-ranking: every model ranks all 11,607 artists, with no sampled-negative shortcuts (which are known to inflate and re-order recommender metrics, Krichene & Rendle, KDD 2020).

How to read this report

Page What it covers
The data The Last.fm-360K core, its sparsity and long tail, and why the small 2k set was a trap.
Methodology The frozen harness, the leakage-safe split, the disciplined search, and how a recommendation is generated.
Results The model leaderboard, significance testing, calibration against published SOTA, and beyond-accuracy metrics.
The pivot Why the served model changed from ALS (on 2k) to EASE (on 360K), the core story.
Limitations An honest account of what the model cannot do, and the ethical considerations.

Why two artifacts? This report is the read-the-work deliverable. The Streamlit app is the play-with-it deliverable, the same model served live, with an interactive recommender and artist-to-artist radio. Both import one shared inference core (src/serving.py); a FastAPI service exposes the same model as a JSON API.

The engineering, in one view

  • Model: EASE (Steck, WWW 2019), a closed-form linear item-item autoencoder, \(B = -P / \operatorname{diag}(P)\) where \(P = (X^\top X + \lambda I)^{-1}\); no hidden layers, no SGD.
  • Data: Last.fm-360K, filtered to a recommendable core of 39,499 users × 11,607 artists (1.68M interactions).
  • Evaluation: frozen eval_core.py, per-user leakage-safe holdout, 5,000-resample paired bootstrap for significance.
  • Serving: one shared src/serving.py behind a Streamlit app and a FastAPI service.
  • Rigor: deterministic given seeds, 137 tests, ruff-clean, CI on every push.