Can you help me choose a recommendation engine for a marketplace with low traffic, new listings, and strict matching rules?

Yes — for a marketplace with low traffic, many new listings, and strict matching rules, the safest choice is usually a hybrid rule-based + content-based recommendation engine, not a pure collaborative filtering system.

Best fit: Hybrid content-based + rules engine

Why it fits your case

• Low traffic: collaborative filtering needs lots of interaction data to work well.
• New listings: content-based methods can recommend items immediately using listing attributes.
• Strict matching rules: a rules layer can enforce hard constraints before ranking.

Recommended architecture

1. Hard-filtering / eligibility layer

Use rules to exclude impossible matches:

• location constraints
• price range
• category compatibility
• availability
• compliance/safety rules
• user preferences that are non-negotiable

This ensures recommendations are valid before scoring.

2. Content-based ranking

Represent each listing and user profile using structured features:

• title, description
• category, brand, attributes
• location
• price
• condition
• tags

Then score candidate listings by similarity to:

• user’s viewed/saved/purchased items
• declared preferences
• search queries

This works well for new items because they don’t need historical interactions.

3. Lightweight personalization

If you have some interaction data, add:

• session-based signals
• implicit feedback boosts
• popularity within a narrow segment
• recency boosts

This gives some personalization without relying on large-scale behavioral data.

4. Optional reranking

After the initial candidate set:

• rerank with business priorities
• diversify results
• avoid duplicates
• promote fresh/new listings

---

Models to consider

Good starting point

• Rules engine + TF-IDF / BM25 + feature weighting
  • simple
  • explainable
  • easy to maintain
  • strong for textual and structured marketplace data

Better if you have enough data

• Two-tower retrieval model or learning-to-rank
  • useful once you have enough interactions
  • can blend content and behavior
  • more complex to deploy

Not ideal as the primary engine

• Pure collaborative filtering
  • struggles with cold start
  • weak for new listings and sparse traffic

---

Practical recommendation

If you’re starting now, I’d choose:

Rules-based filtering + content similarity ranking + popularity/recency rerank

That gives you:

• immediate usefulness for new listings
• compliance with strict constraints
• graceful performance under low traffic
• simpler debugging and control

---

Decision rule of thumb

Choose this if:

• most listings are new
• user activity is sparse
• matching constraints matter a lot
• explainability is important

Consider more advanced ML later if:

• you accumulate enough clicks, saves, messages, and purchases
• you want better personalization beyond similarity and rules

---

If you want, I can also help you with:

1. a specific engine shortlist (e.g. Elasticsearch, OpenSearch, Pinecone, custom ML)
2. a reference architecture
3. a scoring formula
4. a MVP roadmap for the first version of the recommender