Wrap-up & Next Steps

Thank you for participating!

What We Covered Today

A Complete Journey Through Modern RecSys

Part I: Foundations ✅

Classical Recommender Systems

• Collaborative filtering fundamentals
• Matrix factorization and embeddings
• EASE - Embarrassingly Shallow Autoencoders
• Two-tower models for scalable retrieval
• Retrieval → Ranking pipeline

Key Insight: Embeddings are the foundation of all modern recommendation systems

Part II: Sequential Models ✅

Transformers for Recommendations

• Transformer architecture (attention mechanisms)
• Self-Attentive Sequential Recommendation (SASRec)
• Capturing temporal patterns in user behavior
• From static to dynamic user preferences

Key Insight: Sequence matters! Past interactions reveal future intent

Part III: LLM-Powered Recommendation ✅

Generative AI Meets RecSys

• Zero-shot recommendation without training data
• Metadata augmentation with LLMs
• Solving the cold-start problem
• Semantic understanding beyond collaborative signals
• Conversational recommendation

Key Insight: LLMs unlock capabilities impossible with traditional methods

Part IV: Generative Pages ✅

Beyond Ranked Lists

• Re-ranking and slate optimization
• Diversity and coverage constraints
• Whole page generation
• Multi-carousel experiences

Key Insight: Modern platforms generate entire personalized experiences

Key Takeaways

1. Fundamentals still matter - Collaborative filtering, embeddings, and two-tower models remain essential
2. Transformers are everywhere - From NLP to sequential recommendation
3. LLMs are game-changers - Zero-shot capabilities, cold-start solutions, semantic understanding
4. Think beyond lists - Design entire personalized experiences

Take-Home Materials

All materials available on GitHub:

• 📓 Jupyter Notebooks
• 📊 Slide Presentations
• 🐍 Reusable Python utilities

🔗 github.com/jankislinger/recsys-genai

All references: See the References page

Thank You!