I'm building a marketplace and need buyer-seller recommendations

Absolutely — buyer-seller recommendations are a core marketplace feature, and the best design depends on what “recommendation” means for your product.

Common recommendation goals

You might want to recommend:

• Buyers to sellers: “Which sellers should I show this buyer?”
• Sellers to buyers: “Which products/sellers should I show this buyer?”
• Matches: “Which buyer-seller pairs are most likely to transact?”

Typical signals to use

Good marketplace recommendations usually combine several signals:

1) Behavioral signals

• Clicks
• Views
• Saves / favorites
• Add-to-cart
• Purchases
• Message replies
• Search queries

2) Item/seller attributes

• Category
• Price range
• Brand
• Location / shipping region
• Availability
• Ratings / reviews
• Seller response time
• Inventory freshness

3) Buyer attributes

• Budget
• Location
• Past purchases
• Preferred categories
• Brand affinity
• Session intent

4) Marketplace constraints

• Distance / shipping feasibility
• Price compatibility
• Compliance / eligibility
• Seller capacity
• Fraud / trust scores

Recommendation approaches

Start simple

If you’re early-stage, use rules + ranking:

• Filter to eligible sellers
• Rank by relevance score
• Boost sellers with good conversion, fulfillment, and ratings
• Penalize out-of-stock or slow sellers

Then add ML

A strong marketplace system often uses:

• Candidate generation: retrieve a few hundred likely matches
• Ranking model: predict probability of click/purchase/message
• Re-ranking: diversity, fairness, business rules

Common models

• Collaborative filtering
• Content-based matching
• Two-tower retrieval models
• Gradient-boosted ranking models
• Learning-to-rank models
• Graph-based recommendations

A practical scoring formula

For a buyer (b) and seller (s):

[ score(b,s)= w_1 \cdot relevance

• w_2 \cdot purchase_prob
• w_3 \cdot trust
• w_4 \cdot availability

• w_5 \cdot friction ]

Where:

• relevance = category/intent match
• purchase_prob = likelihood of conversion
• trust = ratings, dispute rate, response rate
• availability = in-stock, shipping, active listings
• friction = price mismatch, distance, delay

Cold start handling

For new buyers:

• Popular items/sellers in their category or region
• Onboarding preferences
• Session-based recommendations

For new sellers:

• Use listing content, category, price, location
• Boost until enough performance data exists

Metrics to track

Offline:

• Precision@K
• Recall@K
• NDCG
• MAP

Online:

• CTR
• Conversion rate
• GMV
• Messages sent
• Repeat purchase rate
• Seller response rate
• Cancel/refund rate

Important marketplace-specific concerns

• Two-sided optimization: not just buyer clicks; also seller quality and fairness
• Liquidity: ensure matches are actually transactable
• Exploration vs exploitation: try new sellers too
• Diversity: avoid showing near-duplicates
• Trust & safety: filter bad actors

Recommended MVP

If you want a fast first version:

1. Build a candidate set using filters
2. Score with a weighted heuristic
3. Add popularity and personalization
4. Log all events
5. Train a simple ranking model later

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

1. a recommendation architecture for your marketplace,
2. a ranking formula / scoring model,
3. a data schema for logging events,
4. or example code for buyer-seller matching.