BERT4Rec + RAG Recommender

🎯 Context & Problem

Educational platforms like MOOCs and professional training portals offer thousands of learning objects. The challenge: how do you recommend the right content to the right learner at the right moment, especially when user history is sparse (cold start) or the catalog evolves rapidly?

Classical collaborative filtering fails on cold-start. Content-based filtering ignores sequential learning patterns. This project proposes a hybrid approach that combines:

• BERT4Rec — captures sequential learning patterns from interaction history using a bidirectional Transformer
• RAG — enriches recommendations with semantic retrieval from the content catalog, handling cold-start and providing explainability

This work is the core of the industrial PhD thesis at UCBL/LIRIS in partnership with Inokufu, and has led to publications at WISE 2025, AICCSA 2025 (Best Paper), and EGC 2026.

🏗️ Technical Architecture

The system runs in two phases: offline training of the sequential model, and online hybrid inference combining BERT4Rec predictions with RAG retrieval.

• BERT4Rec trained with masked item prediction on sequential interaction data — capturing "what comes next" in a learning journey
• Category-aware prompt engineering (AICCSA 2025) enriches the RAG prompt with skill/category context for more precise retrieval
• In-Context Learning enables few-shot recommendation without retraining, drastically reducing cold-start issues
• FAISS + ChromaDB serve as the dual retrieval backends — FAISS for speed, ChromaDB for semantic richness
• The full pipeline is exposed via FastAPI with async endpoints for production use

📊 Results & Key Insights

• Category-aware RAG significantly improved recommendation coverage and relevance compared to pure sequential models on real EdTech datasets
• In-Context Learning approach eliminated cold-start issues without requiring additional fine-tuning
• System deployed in production at Inokufu, serving real learners on professional training paths

📄 Related Publications

• WISE 2025 — "CLARE: A Category-Aware RAG-Based Framework for Recommending Learning Objects"
• AICCSA 2025 🏆 Best Paper — "Prompt-Based Recommendation with In-Context Learning and Category-Enriched Modeling"
• EGC 2026 — CLARE (French version, extended)