Multi-Expert Hybrid Retrieval for Movie Recommendation

M.S. Thesis — Learned query routing over semantic, lexical, collaborative, and emotion-aware experts for prompt-conditioned recommendation.

M.S. Thesis · IIT Bombay · Guide: Prof. Arpit Agarwal · Aug 2025 – Present


Overview

Standard recommender systems rely on a single retrieval signal. This thesis argues for a mixture-of-experts approach where four complementary retrieval experts — semantic, lexical, collaborative, and emotion-aware — are dynamically combined based on the query.

Key Contributions

  • Multi-expert architecture: four independent retrieval experts, each capturing a distinct user intent signal
  • Learned query router: trained with a Bradley–Terry–Luce pairwise loss over 9,000 human-judged movie preference pairs; outputs a soft mixture weight per expert per query
  • Evaluation: up to 4.5% higher pairwise agreement over strong single-expert and reranker baselines (BGE-Reranker, Qwen-1.5B, Kimi-K2) on MovieLens-100K

Stack

PyTorch · Hugging Face Transformers · MovieLens-100K · BM25 · FAISS · BTL Ranking Loss