Binge Watch: Reproducible Multimodal Benchmarks Datasets for Large-Scale Movie Recommendation on MovieLens-10M and 20M

Binge Watch: Reproducible Multimodal Benchmarks for MovieLens-10M and 20M Description This repository provides M³L-10M and M³L-20M, large-scale and fully reproducible multimodal extensions of the MovieLens-10M and MovieLens-20M datasets. The datasets augment the original MovieLens interaction data with rich textual, visual, acoustic, and video features extracted from movie plots, posters, and trailers. By following a documented pipeline and avoiding aggressive item filtering, these datasets preserve over 93% of the original rating data while providing high-coverage side information for multimodal recommendation research. Dataset Statistics Metric M³L-10M M³L-20M Total Users 69,878 138,493 Total Items (Full Coverage) 9,031 19,009 Total Ratings 9,409,884 18,777,965 Sparsity 98.66% 99.29% Interaction Coverage 94.10% 93.89% Included Multimodal Features High-dimensional latent representations are provided for each item, extracted using state-of-the-art encoders Textual (Plots): MiniLM, MPNet, and CLIP-Text Visual (Posters): VGG16, Vision Transformer (ViT), and CLIP-Image Acoustic (Trailers): VGGish, Whisper, and AST Video (Trailers): SlowFast (R50), R(2+1)D, and MVIT Resource Contents Interaction Data (MMRec format): Train, Valid, Test data following an 8:1:1 split. This is ready to use in the MMRec framework Check our GitHub repository for the configuration files and the MMRec source code Feature JSONs: Individual mapping files for each encoder (e.g., text_mpnet.json) containing item ID-to-embedding pairs They use the original ML-10M/-20M item IDs Useful to integrate the same features into any other framework Original rating files are available in GroupLens or our repository Raw Data: we also release the raw data files, to foster REPROducibility of our resource Text we collect from IMDB Poster URLs, obtained from IMDB Trailer URLs, obtained from IMDB