Search & Personalisation
Recommendation Engine
Hybrid recommender where every suggestion carries the reason it was made.
The Problem
Cross-sell only works if shoppers see relevant items — and merchandisers will only trust a recommender that can explain itself.
How It Works
- Compute three normalised signals: popularity, content similarity (category, brand, use case, price bucket) and collaborative co-purchase.
- Blend them: 0.2 popularity + 0.4 content + 0.4 collaborative.
- Exclude owned items, rank, and generate an explanation naming the product that drove each result.
Technically Interesting
Classic recommender design with no black box: the top signal and the “driver” product are recorded, so explanations are a by-product of scoring, not an afterthought.
Why It Matters
Explainability is what lets recommendation logic move from experiment to something a business will put on a storefront.
Scope & honesty. Synthetic data (160 customers, 1,102 purchases, 124 products). Classical techniques — no embeddings or LLM.