We outgrew our old recommender and need something better than rules

Absolutely — once a rules-based recommender starts to creak, the next step is usually a machine-learning ranking/recommendation system.

A good modern setup typically looks like this:

1) Start with a clear recommendation goal

What are you optimizing for?

• clicks / CTR
• conversions / revenue
• watch time / retention
• “relevant items” / discovery
• long-term user satisfaction

That choice changes the model and metrics.

2) Use a two-stage architecture

This is the standard scalable pattern:

Candidate generation

• Quickly narrow millions of items down to a few hundred or thousand
• Methods: collaborative filtering, item embeddings, approximate nearest neighbors, retrieval models

Ranking

• Score the candidate set with richer features
• Methods: gradient-boosted trees, deep ranking models, learning-to-rank, wide & deep, two-tower + ranker

3) Build on data, not hand-written rules

Useful signals:

• user-item interactions: clicks, purchases, likes, dwell time, skips
• item metadata: category, price, brand, text, image embeddings
• user context: device, time, locale, session intent
• freshness/popularity signals
• negative feedback / non-clicks

4) Choose a model approach based on maturity

If you want a practical path:

Fastest upgrade from rules

• popularity + personalization
• content-based filtering
• collaborative filtering with matrix factorization

Stronger production system

• candidate retrieval with embeddings
• ranking model with XGBoost/LightGBM
• online re-ranking for diversity/freshness/business rules

Advanced

• deep retrieval + deep ranker
• sequence/session-aware models
• bandits / reinforcement learning for exploration

5) Measure it properly

Offline:

• Precision@K, Recall@K, NDCG@K, MAP
• coverage, diversity, novelty

Online:

• A/B tests on CTR, conversion, revenue, retention
• guardrails: latency, bounce rate, complaint rate, churn

6) Don’t forget the plumbing

You’ll need:

• event tracking and a clean feature store
• training/serving consistency
• model monitoring and retraining
• cold-start handling for new users/items
• fallback logic when data is sparse

7) A pragmatic migration path

If you want to replace rules without a huge rewrite:

1. Keep rules as fallback/guardrails
2. Add personalization to existing ranking
3. Introduce candidate retrieval
4. Train a learning-to-rank model
5. Gradually reduce rule dependence

If you want, I can help you design the recommender for your specific product — for example:

• e-commerce
• media/content feed
• B2B SaaS
• marketplace
• news/articles

If you tell me your domain, data you have, and scale, I can suggest the best architecture and model stack.