New Soundz: Artist Recommendation Engine + Spotify Analytics

machine-learning

recommender-systems

web-app

A hybrid collaborative-filtering recommender that turns your Spotify listening history into personalized artist recommendations, deployed as a live Streamlit web app.

Published

May 7, 2026

Overview

New Soundz is a multi-user web app that connects to your Spotify account, analyzes your listening history, and recommends artists you’d probably like but haven’t heard yet. Built end-to-end as a personal project to ship a real ML system to production rather than just stop at a notebook.

Try the demo live app →

The app is in Spotify’s developer-mode access tier, which caps usage at 5 manually allow-listed users. If you’re a recruiter or visitor and the OAuth login is gated, there’s a built-in demo mode that walks through a simulated session against my own listening data.

Email me @rwquinlan@outlook.com if you would like to be added to the allowed user list!

Tech stack

Python / Streamlit / LensKit / Spotipy / Last.fm API / SQLite / pandas / numpy

How it works

The recommender is a hybrid collaborative-filtering system with three layers:

Collaborative Filtering backbone. Alternating Least Squares (ALS) trained offline on the Last.fm 360K dataset (~360K users × 270K artists, ~17M listening events). Vectorizes every artist into a learned embedding space.
Fold-in inference for new users. When a user logs in with Spotify, their top-artist history is converted into a sparse user vector and folded into the trained ALS model in a single matrix solve; sub-100ms per request, no retraining needed.
Proxy substitution + content fallback. The Last.fm dataset is from 2008–2009, so roughly half of any modern listener’s plays are for artists not in the CF vocabulary (Kendrick, Tyler, Frank Ocean, etc.). For unmatched artists, the system queries Last.fm’s similar-artists API, intersects with the CF vocab, and contributes decayed weights to the user vector. If the result is still too sparse, a content-based scorer ranks candidates by genre-tag overlap.

A separate reverse-proxy expansion step walks the Last.fm similarity graph from the CF recommendations to surface contemporary artists that postdate the training data - the difference between recommending “bands like Radiohead in 2008” and “bands like Radiohead today.”

Two sliders let listeners tune the output: popularity dampening pushes big-name picks out of the way so deeper cuts can surface, and diversity widens the net beyond your usual style to cover adjacent genres.

Repository

rqhq/New_Soundz →

Status

In active development. Day-to-day progress and design decisions are documented in the repo’s project log. Known limitations:

Spotify dev-mode user cap: production access requires Spotify Extension Review (2–6 weeks). Currently allow-listed for 5 users; demo mode covers the rest.
Pre-2010 training data: the Last.fm 360K snapshot ends in 2009, so coverage for newer artists relies on the proxy + reverse-proxy paths via Last.fm .
Genre cache warm-up: first-time recommendations for an unseen artist trigger Last.fm/Spotify API lookups; cached afterward.

What I’d build next

Adding Other Streaming Services: I would love to expand the user-base to other streaming services (i.e. Apple Music, Pandora, Amazon Music). I just chose to stick with Spotify as that is the service I use currently.
Audio Feature Analytics: Spotify’s API used to provide audio feature statistics but they were removed a few years ago. If I could find a way to access them, I’d implement different visualizations and analyses that dive into a user’s audio profile.
Exporting Recommendations: Adding a feature that populates a Spotify playlist with the recommended artists’ tracks. This would allow users to skip the hastle of having to individually explore each recommended artist and would populate say 2-3 songs per artist recommended.

--- title: "New Soundz: Artist Recommendation Engine + Spotify Analytics" description: "A hybrid collaborative-filtering recommender that turns your Spotify listening history into personalized artist recommendations, deployed as a live Streamlit web app." date: "2026-05-07" categories: [machine-learning, recommender-systems, web-app] image: ../images/new-soundz-hero.png image-alt: "Thumbnail for the New Soundz Spotify recommender app" format: html: code-fold: true code-tools: true --- ## Overview New Soundz is a multi-user web app that connects to your Spotify account, analyzes your listening history, and recommends artists you'd probably like but haven't heard yet. Built end-to-end as a personal project to ship a real ML system to production rather than just stop at a notebook. [**Try the demo live app →**](https://new-soundz.streamlit.app/) The app is in Spotify's developer-mode access tier, which caps usage at 5 manually allow-listed users. If you're a recruiter or visitor and the OAuth login is gated, there's a built-in demo mode that walks through a simulated session against my own listening data. Email me @rwquinlan@outlook.com if you would like to be added to the allowed user list! ## Tech stack Python / Streamlit / LensKit / Spotipy / Last.fm API / SQLite / pandas / numpy ## How it works The recommender is a **hybrid collaborative-filtering system** with three layers: 1. **Collaborative Filtering backbone.** Alternating Least Squares (ALS) trained offline on the Last.fm 360K dataset (~360K users × 270K artists, ~17M listening events). Vectorizes every artist into a learned embedding space. 2. **Fold-in inference for new users.** When a user logs in with Spotify, their top-artist history is converted into a sparse user vector and folded into the trained ALS model in a single matrix solve; sub-100ms per request, no retraining needed. 3. **Proxy substitution + content fallback.** The Last.fm dataset is from 2008–2009, so roughly half of any modern listener's plays are for artists not in the CF vocabulary (Kendrick, Tyler, Frank Ocean, etc.). For unmatched artists, the system queries Last.fm's similar-artists API, intersects with the CF vocab, and contributes decayed weights to the user vector. If the result is still too sparse, a content-based scorer ranks candidates by genre-tag overlap. A separate **reverse-proxy expansion** step walks the Last.fm similarity graph from the CF recommendations to surface contemporary artists that postdate the training data - the difference between recommending "bands like Radiohead in 2008" and "bands like Radiohead today." Two sliders let listeners tune the output: **popularity dampening** pushes big-name picks out of the way so deeper cuts can surface, and **diversity** widens the net beyond your usual style to cover adjacent genres. ## Repository [**rqhq/New_Soundz →**](https://github.com/rqhq/New_Soundz) ## Status In active development. Day-to-day progress and design decisions are documented in the repo's project log. Known limitations: - **Spotify dev-mode user cap**: production access requires Spotify Extension Review (2–6 weeks). Currently allow-listed for 5 users; demo mode covers the rest. - **Pre-2010 training data**: the Last.fm 360K snapshot ends in 2009, so coverage for newer artists relies on the proxy + reverse-proxy paths via Last.fm . - **Genre cache warm-up**: first-time recommendations for an unseen artist trigger Last.fm/Spotify API lookups; cached afterward. ## What I'd build next - **Adding Other Streaming Services**: I would love to expand the user-base to other streaming services (i.e. Apple Music, Pandora, Amazon Music). I just chose to stick with Spotify as that is the service I use currently. - **Audio Feature Analytics**: Spotify's API used to provide audio feature statistics but they were removed a few years ago. If I could find a way to access them, I'd implement different visualizations and analyses that dive into a user's audio profile. - **Exporting Recommendations**: Adding a feature that populates a Spotify playlist with the recommended artists' tracks. This would allow users to skip the hastle of having to individually explore each recommended artist and would populate say 2-3 songs per artist recommended.