I'm building a recommendation system for cart add-ons and bundles

Great use case. For cart add-ons and bundles, the recommendation problem is usually best framed as “given the current cart, predict the next best item(s) that increase conversion and basket value without hurting user experience.”

A practical approach:

1) Define recommendation goals

You may want separate models/rules for:

• Add-ons: complementary low-cost items
• Bundles: items commonly bought together
• Upsells: higher-end alternatives
• Cross-sells: related items from different categories

2) Use strong signals

Common inputs:

• Current cart contents
• Product/category embeddings
• Co-purchase history
• Session behavior
• User purchase history
• Price range, brand, compatibility, inventory
• Context: device, time, location, intent

3) Candidate generation

Start with a few ways to produce a shortlist:

• Item-to-item co-occurrence
• Association rules / market basket analysis
• Collaborative filtering
• Embedding similarity
• Category-based heuristics

4) Ranking layer

Rank candidates by a learned score using features like:

• Compatibility with cart
• Predicted conversion probability
• Incremental revenue
• Margin
• Diversity / redundancy
• Stock availability
• Popularity priors

5) Constraints and business rules

Important in cart add-ons:

• Don’t recommend duplicates
• Respect compatibility
• Avoid out-of-stock items
• Limit total number shown
• Filter by price threshold and shipping constraints
• Ensure bundle items actually work together

6) Evaluation

Track:

• CTR on recommendations
• Add-to-cart rate
• Conversion rate
• Average order value
• Revenue per session
• Bundle attach rate
• Long-term retention / return rate

Offline metrics:

• Precision@K / Recall@K
• MAP / NDCG
• Coverage and diversity
• Calibration

7) Simple baseline to start with

A good MVP:

1. For each cart item, retrieve top co-purchased items.
2. Exclude items already in cart.
3. Score by frequency, margin, and category compatibility.
4. Deduplicate across items.
5. Show top 3–5 recommendations.

8) More advanced option

If you have enough data, use:

• Two-tower retrieval model for cart context vs item
• Learning-to-rank model for final ordering
• Sequence model / transformer for cart composition
• Graph-based model for product relationships

If you want, I can help you with any of these next:

1. System architecture
2. Model selection
3. Feature engineering
4. Data schema
5. Evaluation plan
6. Python / SQL example
7. Cold-start strategy