Louis Rollet

Machine Learning / Recommender Systems

Movie Recommendation System

Movie recommendation: from Matrix Factorization to hybrid Neural CF, benchmarked against a zero-shot LLM.

Role
Machine learning project contributor
When
Dec 2025
Status
course project

A Tsinghua machine learning course project comparing classical recommender-system methods, neural collaborative filtering, and a zero-shot LLM baseline for movie recommendation.

The problem

Recommend movies to users from historical preferences while comparing whether specialized recommender models outperform a general-purpose LLM used without task-specific training.

Three recommenders, one question

Does a specialized recommender still beat a general-purpose LLM that has never been trained on the task? The project puts both extremes and a hybrid in the same ring.

Classical baseline

Matrix Factorization

Decomposes the sparse user-movie rating matrix into latent factors, the standard collaborative-filtering reference point.

Main model

Hybrid Neural CF

Neural collaborative filtering combining learned user/item embeddings with interaction history for personalized ranking.

Benchmark

Zero-shot LLM

A general-purpose LLM prompted with a user's history and asked to recommend, with no task-specific training, to test how far prompting alone gets.

The write-up

Louis Rollet et al. "Movie recommendation: from Matrix Factorization to hybrid Neural CF, benchmarked against a zero-shot LLM." Tsinghua University, Machine Learning course, 2025.

OpenReview submission ↗

Features

  • Matrix Factorization baseline
  • Hybrid Neural Collaborative Filtering model
  • Zero-shot LLM benchmark
  • Movie preference modeling from user interaction history

What I did

  • Implemented recommendation approaches based on matrix factorization and collaborative filtering concepts.
  • Built or contributed to a hybrid Neural Collaborative Filtering pipeline for personalized movie recommendations.
  • Benchmarked model behavior against a zero-shot LLM recommendation baseline.
  • Compared approaches from both recommendation quality and implementation complexity perspectives.

Project timeline

Machine-learning course project comparing matrix factorization, hybrid neural collaborative filtering, and a zero-shot LLM recommender baseline.

  1. Dec 2025Idea

    Recommendation problem defined

    Defined the movie-recommendation task from user interaction history such as liked and viewed movies.

  2. Dec 2025Development

    Matrix factorization baseline

    Implemented or evaluated matrix-factorization style collaborative filtering as a classical recommender baseline.

  3. Dec 2025Development

    Hybrid Neural CF model

    Extended the recommender comparison toward a hybrid Neural Collaborative Filtering approach.

  4. Dec 2025Impact / Result

    Zero-shot LLM benchmark

    Benchmarked recommender results against a zero-shot LLM baseline.

  5. Dec 2025Release / Delivery

    Final course project

    Delivered the movie recommendation project for the machine-learning course.

Built with

  • Python
  • Machine Learning
  • Recommendation Systems
  • Matrix Factorization
  • Neural Collaborative Filtering
  • LLM evaluation
  • Collaborative filtering