Cross-sell recommendation model for cart page

For a cart page cross-sell recommendation model, the goal is to suggest a small number of relevant add-on products that increase basket value without distracting the shopper.

1) Define the objective

Typical business goals:

• Increase AOV (average order value)
• Increase attach rate for complementary items
• Maintain or improve conversion rate
• Avoid recommending items that conflict with cart contents

2) Recommendation strategy

A strong cart-page model is usually a hybrid recommender:

A. Item-to-item similarity

Recommend products frequently purchased together with items in the cart.

• Example: cart has a coffee machine → recommend filters, pods, descaler

B. Association / market-basket rules

Use co-occurrence signals from historical orders.

• Metrics: support, confidence, lift
• Good for complementary items

C. Content-based matching

Use product metadata:

• category
• brand
• price band
• compatibility
• attributes like size, color, device model

D. Collaborative filtering

Use user and order interaction patterns:

• “users who bought this also bought…”
• Useful when basket is sparse or product graph is rich

E. Context-aware ranking

Re-rank candidates using cart context:

• price sensitivity
• margin
• inventory
• seasonality
• user segment
• device type

3) Candidate generation

Generate maybe 50–200 candidates from:

• co-purchase graph
• item embeddings
• category rules
• “frequently bought together”
• accessories by parent product
• popular items in user’s region or segment

4) Ranking features

Common features for a ranking model:

• overlap with cart categories
• co-purchase count
• lift/confidence
• item similarity score
• price ratio vs cart value
• margin
• discount relevance
• inventory availability
• brand affinity
• user affinity to category/brand
• novelty/diversity penalty
• compatibility score
• delivery time

5) Model choices

Depending on maturity:

Simple baseline

• Rule-based “frequently bought together”
• Good starting point and easy to explain

Learning-to-rank

• XGBoost / LightGBM ranker
• Train on historical carts with purchased vs not purchased candidates

Deep retrieval + ranking

• Two-tower retrieval model for candidates
• Ranker for final ordering
• Best at scale

6) Training labels

Possible positive labels:

• item added from cart-page recommendation
• item purchased in same session after cart view
• item purchased together in same order

Negative labels:

• shown but not clicked/purchased
• sampled items not purchased with similar carts

Watch out for exposure bias: only items shown can be negatives if you have impression logs.

7) Evaluation

Offline metrics:

• Recall@K
• NDCG@K
• MAP@K
• Hit Rate
• Lift vs baseline

Business metrics:

• attach rate
• incremental revenue
• AOV uplift
• conversion impact
• revenue per impression
• margin per impression

Use A/B testing for final validation.

8) UX recommendations

Cart page usually works best with:

• 2–5 recommendations max
• “Frequently bought together” or “Complete your purchase”
• clear relevance explanation
• avoid competing with the main purchase
• show accessories/consumables first
• enforce diversity and avoid duplicates

9) Practical rules

Good cart-page constraints:

• exclude items already in cart
• exclude out-of-stock items
• exclude low-margin or highly substitutable items if they hurt conversion
• avoid recommending items above a large price threshold unless justified
• cap by category diversity
• prefer accessories, consumables, warranties, add-ons

10) Example architecture

1. Build item co-purchase graph from order history
2. Generate candidates from cart items
3. Add content-based and collaborative candidates
4. Score using ranker
5. Apply business rules
6. Return top N recommendations

If you want, I can also provide:

• a system design for this model
• a feature list + training schema
• a Python/SQL example
• or a LightGBM ranking pipeline for cart-page recommendations.