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

TL;DR

The paper addresses the lack of large-scale, reproducible multimodal benchmarks for movie recommendation by introducing M$^3$L-10M and M$^3$L-20M, which extend MovieLens-10M/20M with textual plots, posters, and trailers. It presents an end-to-end, fully documented pipeline that maps MovieLens items to TMDB, retrieves multimodal data, downloads raw content when possible, encodes features with state-of-the-art models, and releases the resulting embeddings and data mappings. Qualitative analyses using t-SNE and modality fingerprinting reveal complementary, non-redundant signals across text, image, video, and audio, while a preliminary quantitative study with MMRec-based models shows that acoustic features and multimodal combinations can significantly improve recommendation performance. Overall, M$^3$L provides reproducible, scalable multimodal benchmarks to advance MRS research and benchmarking at large scale.

Abstract

With the growing interest in Multimodal Recommender Systems (MRSs), collecting high-quality datasets provided with multimedia side information (text, images, audio, video) has become a fundamental step. However, most of the current literature in the field relies on small- or medium-scale datasets that are either not publicly released or built using undocumented processes. In this paper, we aim to fill this gap by releasing M3L-10M and M3L-20M, two large-scale, reproducible, multimodal datasets for the movie domain, obtained by enriching with multimodal features the popular MovieLens-10M and MovieLens-20M, respectively. By following a fully documented pipeline, we collect movie plots, posters, and trailers, from which textual, visual, acoustic, and video features are extracted using several state-of-the-art encoders. We publicly release mappings to download the original raw data, the extracted features, and the complete datasets in multiple formats, fostering reproducibility and advancing the field of MRSs. In addition, we conduct qualitative and quantitative analyses that showcase our datasets across several perspectives. This work represents a foundational step to ensure reproducibility and replicability in the large-scale, multimodal movie recommendation domain. Our resource can be fully accessed at the following link: https://zenodo.org/records/18499145, while the source code is accessible at https://github.com/giuspillo/M3L_10M_20M.

Figures (3)

• Figure 1: Pipeline to collect and enrich multimodal raw data associated to ML-10M and ML-20M.
• Figure 2: t-SNE projections revealing structural differences across modalities. Text forms densely concentrated clusters organized by narrative similarity, where Psycho and Awakenings are positioned in overlapping central regions due to shared psychological themes. Image exhibits spatial separation with lobes defined by visual archetypes. Here Psycho and Awakenings occupy distinct regions reflecting their different color palettes and cinematographic styles. Video and Audio display greater dispersal, capturing kinetic and atmospheric properties: Psycho and Awakenings co-locate in Video space due to shared visual pacing, demonstrating how stylistic modalities encode non-narrative cinematic properties that transcend genre and era.
• Figure 3: Modality Fingerprinting: Comparison of thematic versus stylistic profiles across normalized dimensions.