The Challenge
Traditional recommendation systems at QVC relied on collaborative filtering and basic content matching. These approaches couldn't capture the rich semantic relationships between products, categories, shows, and customer behaviors that define the live retail experience. We needed a recommendation engine that understood context - not just "people who bought X also bought Y," but the deeper why behind purchasing patterns.
My Approach
I architected a graph-based recommendation system using Neo4j as the knowledge graph backbone, leveraging the Graph Data Science (GDS) library and APOC procedures for advanced graph algorithms.
Knowledge Graph Design
The product knowledge graph models:
- Products with enriched attributes (category, brand, show context, seasonal relevance)
- Customers with behavioral profiles built from purchase and engagement history
- Relationships capturing semantic connections - purchases, views, complementary products, show associations, category hierarchies
- Temporal patterns reflecting seasonality and recency
Graph ML Techniques
Using Neo4j GDS, the system applies:
- Node embeddings (GraphSAGE, Node2Vec) for dense product and customer representations
- Community detection for natural customer segmentation
- Link prediction for recommendation candidate generation
- Cypher-based feature engineering feeding downstream ML models
Hybrid Architecture
The graph-based signals are combined with traditional ML approaches - the knowledge graph provides structural and semantic features that enrich models trained on behavioral data. This hybrid approach delivers recommendations that are both accurate and explainable.
Key Decisions & Trade-offs
Graph over pure embeddings: While vector similarity is simpler, the graph gives us explainable recommendation paths - we can trace exactly why a product was recommended. This transparency was critical for merchandising team adoption.
GDS for in-graph computation: Running ML algorithms directly within Neo4j (via GDS) rather than exporting data avoids the latency and complexity of graph-to-tabular pipelines.
Impact
The knowledge graph became a strategic asset beyond recommendations - it serves merchandising planning, content strategy, and customer insights. The graph-based approach captures relationships that flat data structures fundamentally miss.