Fidelio, pipeline hybrid movie recommender system

A pipeline movie recommendation system developed as part of the Machine Learning (ML) course.

Pipeline Architecture

This project implements an hybrid methodology to solve the cold-start problem and improve recommendation accuracy:

🔹 Pipeline: The Gradient Boosted Hybrid

A feature-based machine learning approach that treats recommendation as a regression/ranking problem:

1. Content Encoding: Metadata is encoded into numerical features.
2. Collaborative Features: User and Item embeddings/vectors are used as input features.
3. Gradient Boosting Machine (GBM): A Tree-based model (Hist Gradient Boosting) is trained on these combined features to predict the specific user rating.

Data Sources & Acknowledgments

This project utilizes a custom dataset constructed by merging user ratings with rich movie metadata collected from multiple sources:

• MovieLens Dataset
  User–item ratings are derived from the MovieLens datasets provided by GroupLens Research.
  These ratings form the backbone of the collaborative filtering pipeline.

  • https://grouplens.org/datasets/movielens/

• The Movie Database (TMDB) API
  Rich movie metadata (TMDB IDs, cast, crew, directors, genres) was collected programmatically via the TMDB API and used primarily for content-based modeling.

  • https://developers.themoviedb.org/3

• Rotten Tomatoes Reviews Dataset
  A large-scale dataset of critic and audience reviews scraped from Rotten Tomatoes, used to enrich item representations and provide additional signals for cold-start movies and sentiment-aware modeling.

  • https://www.kaggle.com/datasets/priyamchoksi/rotten-tomato-movie-reviews-1-44m-rows

Trained Model

Due to GitHub file size limitations, the trained models and datasets are not included in this repository.

You can download the trained models and datasets from Hugging Face: https://huggingface.co/jacob14047/Fidelio/