AI-Powered Music Compatibility Matching
Find your musical soulmate using Graph Neural Networks and vector similarity search
Demo • Docs • Report Bug • Request Feature
VibeMatch connects users with similar music taste by analyzing their listening patterns through deep learning. Unlike traditional systems that compare artist lists, VibeMatch captures complex musical relationships through a Graph Neural Network trained on millions of music interactions.
Key Innovation: We don't just match artists you both like—we understand the musical relationships between artists to find truly compatible taste profiles.
- Deep Music Understanding: Graph Neural Network analyzes 338K+ tracks and 6.9K+ artists to learn musical relationships
- Instant Matching: Sub-10ms vector search across user profiles using Qdrant
- Multi-temporal Analysis: Combines long-term preferences with recent listening trends
- Cold Start Solution: Synthetic profiles ensure immediate matches for new users
- Privacy-First: Only uses public Last.fm data, GDPR compliant
graph LR
A[Last.fm Profile] --> B[Multi-Period Fetch]
B --> C[Weighted Embedding]
C --> D[Qdrant Search]
D --> E[Ranked Matches]
style A fill:#e0f2fe,stroke:#0ea5e9,stroke-width:2px,color:#0c4a6e
style B fill:#ffe4e6,stroke:#f43f5e,stroke-width:2px,color:#881337
style C fill:#dcfce7,stroke:#22c55e,stroke-width:2px,color:#14532d
style D fill:#f3e8ff,stroke:#a855f7,stroke-width:2px,color:#581c87
style E fill:#fef3c7,stroke:#eab308,stroke-width:2px,color:#713f12
Step-by-step:
- Authentication: Connect your Last.fm account via OAuth
- Data Fetching: Retrieve listening history across multiple time periods (all-time, 6mo, 3mo, recent)
- Embedding Generation: Your music taste is encoded into a 128D vector using the trained GNN (~500ms)
- Vector Search: Qdrant finds top-K compatible users via cosine similarity (<10ms)
- Results: View matches with compatibility scores and shared artists
graph LR
subgraph Frontend["Frontend Layer"]
direction TB
A[Next.js 15<br/>TypeScript]
A1[Tailwind UI]
A2[Zustand State]
A --> A1
A --> A2
end
subgraph Backend["API Layer"]
direction TB
B[FastAPI<br/>Async Server]
C[Embedding<br/>Service]
D[LRU Cache<br/>8K entries]
B --> C
C --> D
end
subgraph ML["ML Pipeline"]
direction TB
E[PyTorch +<br/>PyG]
F[LightGCN<br/>3 Layers]
G[Embeddings<br/>338K tracks<br/>6.9K artists]
E --> F
F --> G
end
subgraph VectorDB["Vector Search"]
direction TB
H[(Qdrant<br/>Database)]
I[HNSW Index<br/>Cosine Sim]
H --> I
end
subgraph External["External APIs"]
J[Last.fm<br/>OAuth + Data]
end
Frontend -->|REST API| Backend
Backend -->|Vector Query| VectorDB
Backend -->|Auth + Fetch| External
Backend -.->|Embedding Lookup| ML
VectorDB -.->|Store Vectors| Backend
style Frontend fill:#e0f2fe,stroke:#0ea5e9,stroke-width:2px,color:#0c4a6e
style Backend fill:#dcfce7,stroke:#22c55e,stroke-width:2px,color:#14532d
style ML fill:#ffe4e6,stroke:#f43f5e,stroke-width:2px,color:#881337
style VectorDB fill:#f3e8ff,stroke:#a855f7,stroke-width:2px,color:#581c87
style External fill:#fef3c7,stroke:#eab308,stroke-width:2px,color:#713f12
style A fill:#bae6fd,stroke:#0284c7,stroke-width:2px,color:#075985
style B fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
style E fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style H fill:#e9d5ff,stroke:#9333ea,stroke-width:2px,color:#6b21a8
style J fill:#fef08a,stroke:#ca8a04,stroke-width:2px,color:#854d0e
LightGCN (Light Graph Convolutional Network)
graph LR
subgraph GraphStructure["Graph Structure"]
direction TB
T1[Track 1]
T2[Track 2]
T3[Track 3]
A1[Artist 1]
A2[Artist 2]
T1 -.->|authored by| A1
T2 -.->|authored by| A1
T3 -.->|authored by| A2
T1 ---|co-occurrence| T2
A1 ---|similar| A2
end
subgraph Training["GNN Training Pipeline"]
direction TB
G[345K Nodes<br/>2.7M Edges] --> L[3-Layer LightGCN]
L --> E[128D Embeddings]
E --> N[L2 Normalization]
end
subgraph Optimization["Training & Optimization"]
direction TB
N --> BPR[BPR Loss]
BPR --> Adam[Adam Optimizer]
Adam --> M[Model Weights]
end
GraphStructure --> Training
Training --> Optimization
style GraphStructure fill:#e0f2fe,stroke:#0ea5e9,stroke-width:2px,color:#0c4a6e
style Training fill:#ffe4e6,stroke:#f43f5e,stroke-width:2px,color:#881337
style Optimization fill:#dcfce7,stroke:#22c55e,stroke-width:2px,color:#14532d
style G fill:#bae6fd,stroke:#0284c7,stroke-width:2px,color:#075985
style L fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style E fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style N fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style BPR fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
style Adam fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
style M fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
Model Specs:
- Embeddings: 128 dimensions, L2 normalized
- Architecture: 3-layer graph convolution
- Training: Bayesian Personalized Ranking (BPR) loss
- Performance: Recall@10: 0.64, Precision: 1.00
- Graph: 345K nodes (338K tracks, 6.9K artists), 2.7M edges
Model Specs:
- Embeddings: 128 dimensions, L2 normalized
- Architecture: 3-layer graph convolution
- Training: Bayesian Personalized Ranking (BPR) loss
- Performance: Recall@10: 0.64, Precision: 1.00
- Graph: 345K nodes (338K tracks, 6.9K artists), 2.7M edges
|
Frontend
|
Backend
|
| Metric | Value |
|---|---|
| Embedding Generation | ~500ms |
| Vector Search | <10ms |
| End-to-End Latency | <800ms |
| Model Size | ~168MB |
| User Embedding Coverage | ~95% (exact + fuzzy + zero-shot) |
Built from Last.fm data:
- Source: Last.fm augmented dataset with artist similarity graph
- Tracks: 338,046 Last.fm tracks
- Artists: 6,899 unique artists
- Relationships: 2.7M edges (track-artist, track-track, artist-artist)
- Genre coverage: 95.9% of tracks have genre assignments
Coverage breakdown for user embeddings:
- Exact matches: ~60%
- Fuzzy matches: ~25%
- Zero-shot inference: ~10%
- Missing: <5%
Multi-temporal weighted average with consistency boosting:
graph LR
subgraph Sources["Data Sources"]
direction TB
A1[Overall<br/>45%]
A2[6 Months<br/>25%]
A3[3 Months<br/>15%]
A4[Recent 200<br/>15%]
end
subgraph Lookup["Lookup Strategy"]
direction TB
B1[Exact Match<br/>O1 - 60%]
B2[Fuzzy FAISS<br/>~25%]
B3[Zero-Shot<br/>~10%]
B4[LRU Cache<br/>8K entries]
end
subgraph Weighting["Weighting Pipeline"]
direction TB
C1[Consistency Boost<br/>Multi-period × 1.4]
C2[Temporal Decay<br/>0.5^days/30]
C3[Playcount Weight<br/>log1p]
end
subgraph Output["Output"]
direction TB
D[128D Normalized<br/>User Vector]
end
Sources --> Lookup
Lookup --> Weighting
Weighting --> Output
B4 -.->|Cache Hit| B1
style Sources fill:#e0f2fe,stroke:#0ea5e9,stroke-width:2px,color:#0c4a6e
style Lookup fill:#ffe4e6,stroke:#f43f5e,stroke-width:2px,color:#881337
style Weighting fill:#dcfce7,stroke:#22c55e,stroke-width:2px,color:#14532d
style Output fill:#f3e8ff,stroke:#a855f7,stroke-width:2px,color:#581c87
style A1 fill:#bae6fd,stroke:#0284c7,stroke-width:2px,color:#075985
style A2 fill:#bae6fd,stroke:#0284c7,stroke-width:2px,color:#075985
style A3 fill:#bae6fd,stroke:#0284c7,stroke-width:2px,color:#075985
style A4 fill:#bae6fd,stroke:#0284c7,stroke-width:2px,color:#075985
style B1 fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style B2 fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style B3 fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style B4 fill:#fecdd3,stroke:#e11d48,stroke-width:2px,color:#9f1239
style C1 fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
style C2 fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
style C3 fill:#bbf7d0,stroke:#16a34a,stroke-width:2px,color:#166534
style D fill:#e9d5ff,stroke:#9333ea,stroke-width:2px,color:#6b21a8
Fallback Hierarchy:
- Exact match → Precomputed embedding (O(1)) - ~60% coverage
- Fuzzy match → FAISS similarity search - ~25% coverage
- Zero-shot → Weighted average of similar artists - ~10% coverage
- LRU Cache → 8K entries, ~75% hit rate
Fallback Hierarchy:
- Exact match → Precomputed embedding (O(1)) - ~60% coverage
- Fuzzy match → FAISS similarity search - ~25% coverage
- Zero-shot → Weighted average of similar artists - ~10% coverage
- LRU Cache → 8K entries, ~75% hit rate
Graph Neural Networks combine collaborative filtering with content-based features and relational structure. Each GNN layer aggregates neighbor information:
e_i^(k+1) = Σ(neighbors) [1/√(degree_i × degree_j)] × e_j^(k)
Final embedding averages all layers (local + global context).
References
- He et al. (2020) - "LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation"
- Rendle et al. (2009) - "BPR: Bayesian Personalized Ranking from Implicit Feedback"
- Only public Last.fm API data
- No scraping, respects rate limits
- OAuth 1.0, no password storage
- Anonymous embeddings (non-reversible)
- GDPR compliant with right to deletion
MIT License