A Content-Driven Micro-Video Recommendation Dataset at Scale

Abstract

Micro-form videos have emerged as a popular form of content, leading to extensive research in micro-video recommendation with significant implications for the entertainment, advertising, and e-commerce industries. However, the lack of publicly available large-scale micro-video datasets presents a challenge for developing effective recommender systems. To address this challenge, we introduce a comprehensive and diverse micro-video recommendation dataset, referred to as ''MicroLens.'' This dataset comprises nine million user-item interaction behaviors, one million users, and 91 thousand full-length micro-videos. It includes rich modality information such as titles, cover images, and audio associated with the videos. MicroLens serves as a benchmark for the content-driven micro-video recommendation, allowing researchers to leverage diverse video modality information, particularly the raw video features, to enhance the effectiveness of recommender systems. This goes beyond the traditional reliance on item IDs or off-the-shelf pre-extracted video/visual features, providing new avenues for improving recommendation accuracy and personalization. We have conducted extensive experiments on MicroLens, benchmarking multiple recommender models and video encoders, which have provided valuable insights into the performance of micro-video recommendation. We anticipate that this dataset will not only benefit the recommender system community but also foster advancements in the field of video understanding. Our datasets, code, and additional documents are available at https://github.com/westlake-repl/MicroLens.