Design a Spotify Recommendation System

Design a machine learning system to recommend songs and playlists on Spotify.

Requirements

Functional:

• Recommend songs for "Discover Weekly" playlist
• Recommend songs for radio stations
• Real-time "next song" recommendations
• Personalized home page recommendations

Non-functional:

• Generate Discover Weekly for 400M+ users weekly
• Real-time recommendations < 100ms
• Handle catalog of 80M+ songs

Metrics

Offline Metrics

Metric	Description
Recall@K	Fraction of user's future listens in top K
NDCG	Ranking quality of recommendations
Coverage	Percentage of catalog recommended
Diversity	Variety in recommendations

Online Metrics

Metric	Description
Stream Rate	Songs played / Songs recommended
Skip Rate	Songs skipped within 30 seconds
Save Rate	Songs added to library
Listen Time	Total listening duration

Business Metrics

• Monthly Active Users (MAU)
• Premium conversion rate
• Session duration
• Artist discovery rate

Architecture

Loading diagram...

Candidate Generation

Approach 1: Collaborative Filtering

Users who like similar songs will like similar new songs.

Matrix Factorization: Decompose user-song interaction matrix R into user factors (U) and song factors (V).

To recommend for user i:

1. Get user embedding U[i] (128-dim vector)
2. Compute dot product with all song embeddings V
3. Return top-N highest scoring songs

Approach 2: Content-Based Filtering

Recommend songs similar to what user already likes.

Audio Features:

• Tempo, key, loudness
• Danceability, energy, valence
• Audio embeddings from neural networks

Metadata Features:

• Genre, artist, album
• Release year
• Lyrics embeddings

Approach 3: Two-Tower Model

Neural network approach for candidate retrieval.

Loading diagram...

Feature Engineering

User Features

Feature	Description
listening_history	Sequence of recent tracks
top_artists	Most listened artists
top_genres	Most listened genres
listening_time_distribution	When they listen
skip_rate	How often they skip
playlist_creation_behavior	Playlists they've made
premium_status	Subscription tier

Song Features

Feature	Description
audio_features	Spotify's audio analysis (tempo, energy, etc.)
audio_embedding	Neural network embedding of audio
artist_embedding	Artist representation
genre_embedding	Genre representation
popularity	Global and regional popularity
release_date	When released
lyrics_embedding	NLP embedding of lyrics

Context Features

Feature	Description
time_of_day	Morning, afternoon, evening, night
day_of_week	Weekday vs weekend
device_type	Phone, desktop, smart speaker
activity_context	Workout, focus, party (if available)

Model Architecture

Discover Weekly Pipeline

Weekly batch job to generate personalized playlists.

Step	Description
1	Candidate Generation: 1000 songs from CF neighbors, content-based similar songs, trending in user's genres
2	Ranking Model: Predict P(stream), P(save), P(skip). Combine into final score.
3	Diversity Optimization: Ensure artist diversity, mix familiar and new, genre balance
4	Quality Filters: Remove explicit if preference set, remove recently played, remove disliked artists

Real-time Radio Recommendations

Loading diagram...

Training

Data Collection

Signal Type	Events	Interpretation
Positive	Stream complete (over 30s)	User engaged with song
	Save to library	Strong preference
	Add to playlist	Curated preference
	Repeat listen	Very strong signal
Negative	Skip early (under 30s)	User didn't like
	Remove from playlist	Changed preference
	Hide song	Explicit dislike

Loss Function

Multi-task learning with weighted losses:

Task	Weight	Rationale
Stream prediction	w1	Primary engagement signal
Save prediction	w2	Strong preference indicator
Skip prediction	w3	Negative signal
Playlist add prediction	w4	Curated preference

Total Loss = w1 x stream_loss + w2 x save_loss + w3 x skip_loss + w4 x playlist_loss

Handling Implicit Feedback

Listening data is implicit - not playing doesn't mean dislike.

Solutions:

• Weighted matrix factorization
• Bayesian personalized ranking (BPR)
• Negative sampling strategies

Cold Start Problem

New Users

Approach	Description
Onboarding survey	Ask for favorite artists/genres
Popular items	Recommend globally popular songs
Demographic similarity	Use age, location-based recommendations
Quick learning	Rapidly update from first few interactions

New Songs

Approach	Description
Content-based	Use audio features and metadata
Artist fans	Recommend to fans of the artist
Editorial playlists	Human curation for initial exposure
Exploration budget	Allocate slots for new content

Serving Infrastructure

Embedding Index

Use Approximate Nearest Neighbor (ANN) search for fast similarity lookup:

Component	Configuration	Purpose
Index Type	Annoy / HNSW / Faiss	Trade-off: build time vs. query speed
Embedding Dim	128	Balance expressiveness vs. storage
Distance Metric	Angular (cosine)	Normalized similarity
Index Trees	100	More trees = better recall, slower
Query K	100-200	Candidates for ranking stage

Query Flow:

1. User embedding -> ANN index query
2. Return top-K most similar song embeddings
3. Latency target: under 10ms for 100 results from 50M songs

Caching Strategy

Data	Cache TTL	Update Frequency
User Embedding	1 hour	Re-compute on significant activity
Song Embeddings	Permanent	Updated daily batch job
Candidate Pool	15 minutes	Per-user, invalidated on context change
Ranking Scores	5 minutes	Short-lived, context-dependent

Monitoring

Quality Metrics

Metric	Formula	Target
Stream Rate	Streams / Impressions	Over 60%
Skip Rate	Early Skips / Streams	Under 25%
Save Rate	Saves / Streams	Over 5%
Discovery Rate	New Artist Streams / Total	15-30%
Diversity Score	Unique Artists / Total Streams	Over 40%

A/B Testing

• Test new models against production
• Segment by user type (new, casual, power users)
• Monitor for novelty effects

Reference

Topic	Description
Exploitation vs exploration	Playing safe favorites keeps satisfaction stable. Discovery keeps the experience fresh.
Personalization depth	Too personalized feels stale. Too random feels irrelevant.
Computation trade-off	Real-time is fresher but more expensive. Batch is efficient but stale.
Popularity bias	Popular songs are safe bets. Long-tail discovery differentiates the platform.