PaperlayerView exact revision on arXiv

Source-linked AI summary

A Content-Driven Micro-Video Recommendation Dataset at Scale

Yongxin Ni, Yu Cheng, Xiangyan Liu, Junchen Fu, Youhua Li, Xiangnan He, Yongfeng Zhang, Fajie Yuan

arXiv:2309.15379v1cs.IR

TL;DR

Large-scale public micro-video datasets with diverse raw content and user behavior are lacking, despite growing interest in micro-video recommendation. MicroLens provides interactions and raw modalities, and benchmarks recommender models and video encoders. VideoRec achieves the highest recommendation accuracy among the compared models, while the study identifies a gap between video understanding and recommendation.

  • Problem

    Public micro-video datasets lack diverse, high-quality video content and user behavior, while existing datasets may offer only IDs or pre-extracted thumbnail features.

    Source: §1 · p. 1

  • Method

    MicroLens provides large-scale user-item interactions and raw video modalities, alongside a benchmark of recommender models and video encoders.

    Source: §1 · p. 2

  • Results

    VideoRec achieves the highest recommendation accuracy among compared models, and the study finds a gap between current video-understanding technologies and video recommendation.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    MicroLens may serve as a pre-training dataset for visually relevant recommendation, and video recommendation as a downstream task may connect video understanding and recommender systems.

    Source: §5 · pp. 9–10

    MicroLens includes public user behaviors for privacy protection, and videos are accessed through URLs and a download tool to avoid copyright issues.

    Source: §2.2 · p. 3

Abstract

from arXiv · show

Micro-videos have recently gained immense popularity, sparking critical research in micro-video recommendation with significant implications for the entertainment, advertising, and e-commerce industries. However, the lack of large-scale public micro-video datasets poses a major challenge for developing effective recommender systems. To address this challenge, we introduce a very large micro-video recommendation dataset, named "MicroLens", consisting of one billion user-item interaction behaviors, 34 million users, and one million micro-videos. This dataset also contains various raw modality information about videos, including titles, cover images, audio, and full-length videos. MicroLens serves as a benchmark for content-driven micro-video recommendation, enabling researchers to utilize various modalities of video information for recommendation, rather than relying solely on item IDs or off-the-shelf video features extracted from a pre-trained network. Our benchmarking of multiple recommender models and video encoders on MicroLens has yielded valuable insights into the performance of micro-video recommendation. We believe that this dataset will not only benefit the recommender system community but also promote the development of the video understanding field. Our datasets and code are available at https://github.com/westlake-repl/MicroLens.

1 Introduction

Micro-video recommendation lacks large-scale public datasets with diverse content and user behavior, limiting work on content-driven models. MicroLens combines large-scale interactions with raw video modalities and benchmarks recommender models and video encoders.

1 Introduction

  • Existing public datasets lack diverse, high-quality micro-video content and user behavior, while MovieLens focuses mainly on longer movie-type videos.Tenrec and KuaiRec provide IDs or pre-extracted thumbnail features, limiting direct learning from raw video content.Source: §1 · pp. 1–2
  • MicroLens contains one billion user-item interactions, 34 million users, and one million micro-videos with original title, cover-image, audio, and video modalities.The dataset is presented as a resource for content-driven recommendation rather than relying only on IDs or pre-extracted features.Source: §1 · p. 2
  • End-to-end models achieve superior performance despite higher computational expense, while the benchmark covers over 10 recommender models and video encoders.The study also reports a gap between current video-understanding technologies and video recommendation, motivating specialized research.Source: §1 · p. 2
  • The empirical study finds a significant gap between current video-understanding technologies and video recommendation, highlighting the need for specialized research.It also explores potential relationships between video understanding and recommender systems.Source: §1 · p. 2

2 MicroLens

MicroLens was assembled from a year of public-platform micro-video data, combining video content and user comments. The paper describes its privacy and access provisions, dataset statistics, and differences from existing recommendation datasets.

2.1 Dataset Construction

  • The collection began with 400,000 popular seed videos and expanded through related-video links to approximately 5 million videos before filtering.Seed videos had more than 10,000 likes; titles, covers, audio, and raw video were retained.Source: §2.1 · pp. 2–3
  • Public comments serve as positive interaction feedback; duplicate comments by a user were removed, and up to 5,000 comments per video were collected.User and video IDs were anonymized.Source: §2.1 · p. 3

2.2 Privacy and Copyrights

  • MicroLens includes only public user behaviors, anonymizes user and item IDs, and provides video URLs plus a tool to access and download videos.The authors describe the original dataset as provided with reference to the ImageNet license.Source: §2.2 · p. 3

2.3 Dataset Analysis

  • MicroLens-100K has a long-tail item-popularity distribution, most users have 5–15 interactions, and most videos are under 400 seconds.The paper contrasts MicroLens with Reasoner, which offers five frames per video and has about 10,000 times fewer users and user-video behaviors.Source: §2.3 · p. 4, §2.4 · p. 5
  • The original dataset contains 34,492,051 users, 1,142,528 items, and 1,006,528,709 interactions; MicroLens-100K is the default evaluation subset.MicroLens-100K contains 100,000 users, 19,738 items, and 719,405 interactions, with 99.96% sparsity.Source: §2.3 · p. 4

2.4 Comparison to Existing Datasets

  • The authors report no existing micro-video recommendation dataset with original video content, while Reasoner is smaller and MovieLens covers movie categories.They also note that MovieLens interactions come from a simulated rating site rather than actual watching behavior.Source: §2.4 · p. 5

3 Experiments Setup

The experiments focus on micro-video recommendation, using MicroLens-100K for the default evaluation. They compare ID-based, feature-augmented, and end-to-end video-encoder approaches under a leave-one-out evaluation setup.

3 Experiments Setup

  • MicroLens-100K is the default evaluation dataset; results on MicroLens-1M are reported in the Appendix.The paper's stated experimental focus is micro-video recommendation.Source: §3 · p. 5

3.1 Baselines and Evaluation

  • The benchmark compares IDRec, which uses item IDs, VIDRec, which adds video features, and VideoRec, which replaces item IDs with an end-to-end learnable video encoder.VideoRec optimizes the recommender and encoder jointly, but has high training costs.Source: §3.1 · p. 5
  • Evaluation uses leave-one-out splits, recent sequences capped at 13 comments, and HR@10, HR@20, NDCG@10, and NDCG@20.The last interaction is test data and the penultimate interaction is validation data.Source: §3.1 · p. 5

3.2 Hyper-parameter Tuning

  • IDRec hyperparameters were searched first; VideoRec requires 10–50 times more compute and training time, so only its top few encoder layers were optimized.VIDRec and VideoRec initially use IDRec's selected hyperparameters, followed by basic searches around them.Source: §3.2 · pp. 5–6

4 Experimental Results

VideoRec achieves the strongest recommendation accuracy among the compared approaches, while the experiments show that video-classification performance and recommendation performance do not always align. The study also finds that video encoders benefit from recommendation-task fine-tuning, and that VideoRec can outperform IDRec for warm items.

4.1 Benchmark Results & Analysis

  • End-to-end VideoRec achieves the highest recommendation accuracy among the compared models, while pre-extracted video features do not necessarily improve over IDRec.The authors attribute VideoRec’s advantage to learning from raw video and collaborative signals together.Source: §4.1 · pp. 6–7

4.2 Video Understanding Meets Recommender Systems

  • Pre-trained video-classification weights improve SASRecV for some encoders, but stronger classification accuracy does not necessarily mean stronger recommendation accuracy.VideoMAE leads video classification without guaranteeing the highest recommendation accuracy; R3D-r18 shows the reverse pattern.Source: §4.2 · pp. 7–8
  • Fine-tuning the top few encoder layers generally outperforms frozen features, yields about a 2-fold improvement over freezing, and can outperform full fine-tuning.Adding DNN layers to frozen features improves accuracy but still falls short of using a fine-tuned encoder.Source: Figure 4 · p. 7, §4.2 · p. 8, Figure 5 · p. 8

4.3 Additional Exploration of VideoRec

  • VideoRec generally outperforms ImageRec when image encoders represent videos using only their cover images.The comparison uses ResNet, Swin Transformer, and MAE image encoders.Source: §4.3 · p. 9, Table 3 · p. 9

Q2: Can VideoRec compete with IDRec in recommending highly popular items?

  • VideoRec using SASRec consistently outperforms IDRec even for very warm items, and the authors report potential replacement of ID features in warm and cold settings.They note that gains are more visible on cold items.Source: Q2: Can VideoRec compete with IDRec in… · p. 9

5 Conclusions and Broader Impact

The paper presents MicroLens as a large, diverse micro-video dataset with rich modalities and benchmarks recommender baselines, including VideoRec. It anticipates uses in multimodal recommendation and visually relevant pre-training, while describing video recommendation as a potential link between video understanding and recommendation research.

5 Conclusions and Broader Impact

  • MicroLens combines text descriptions, images, audio, and raw video with extensive recommender benchmarking, and VideoRec achieves the highest accuracy among compared models.The authors anticipate the dataset will support multimodal and micro-video recommendation research.Source: §5 · p. 9
  • MicroLens may serve as a pre-training resource for visually relevant recommendation, with each video potentially yielding hundreds of images and trillion-level user-image interactions.The paper identifies the lack of large-scale, diverse, high-quality multimodal recommendation datasets as a challenge.Source: §5 · p. 9
  • The authors describe video recommendation as a potential downstream task for video understanding that could foster mutual development between the communities.They characterize fine-grained recommendation using video understanding as a promising direction.Source: §5 · p. 10

A Technical Details for Data Integration

The data integration process distributes download links among multiple download nodes and consolidates the collected media into a root directory. The integration node stores the downloaded data for later analysis.

A Technical Details for Data Integration

B Related Datasets

The supplied passages identify a dataset-comparison table and define a baseline hyperparameter used in video-encoder fine-tuning.

B Related Datasets

  • Table 4 compares datasets and distinguishes pre-extracted visual features from raw image pixels and original full-length audio and video.Source: Table 4 · p. 16

C Hyper-parameter Settings for Baselines

E Details of the Applied Image Encoders

This section records implementation details for image and video encoders, including baseline hyperparameters, video-encoder performance, and image-encoder architectures.

E Details of the Applied Image Encoders

  • Table 5 lists the hyperparameter settings used for the baselines.Source: Table 5 · p. 17
  • Table 6 reports VideoRec performance across 15 video encoders and gives their pretraining frame lengths and sample rates.It also reports ACC@5 for video classification.Source: Table 6 · p. 17
  • Table 7 describes image-encoder network architectures, parameter sizes, and download URLs.Its architecture labels include transformer blocks, attention heads, channel number, and layers per block.Source: Table 7 · p. 17

G Baseline Evaluation and Warm-up Recommendation on MicroLens-1M

The MicroLens-1M appendix reports baseline and warm-up evaluations, with trends consistent with MicroLens-100K, and includes a cold-start comparison organized by item popularity.

G Baseline Evaluation and Warm-up Recommendation on MicroLens-1M

  • The MicroLens-1M baseline and warm-up recommendation trends are consistent with those on MicroLens-100K.Some baselines could not be trained on MicroLens-1M because of excessive GPU memory consumption, so their results are not reported.Source: G Baseline Evaluation and Warm-up… · p. 18
  • Table 10 compares VideoRec and IDRec in regular and warm-start settings using SASRec, with Warm-20 excluding items with fewer than 20 interactions.Source: Table 10 · p. 18

H Recommendation in Cold-start Scenarios

  • Figure 6 plots VideoRec-to-IDRec relative improvement in HR@10 across item groups ordered by popularity; larger group numbers indicate more popular items.Source: Figure 6 · p. 18
  • Table 11 reports recommendation results with side features on MicroLens-100K.Source: Table 11 · p. 19

I Recommendation with Side Features

The study tests popularity and tag categories as side features for SASRecID and finds that they do not clearly improve performance.

I Recommendation with Side Features

  • Popularity is divided into 10 uniform bins, from the top 10% most popular items to the bottom 10%, and tags are treated as categorical features.Source: I Recommendation with Side Features · p. 19
  • The SASRecID experiments compare ID alone with Pop, Tag, and Pop+Tag combinations, where combined features are summed and averaged.Source: I Recommendation with Side Features · p. 19
  • Popularity-level and tag-category side features did not clearly improve recommendation performance.The paper suggests trained item-ID embeddings may already capture latent similarity, popularity, and category information; in very cold-start settings, other features are necessary because item IDs are weak.Source: I Recommendation with Side Features · p. 19

J Comparison between Textual Features and Video Content

The comparison finds that text features perform similarly to item IDs, while the authors point to the greater information in full videos; the collected platform presents a continuous, swipe- or scroll-driven video stream.

J Comparison between Textual Features and Video Content

  • Table 12 compares item-ID, textual-feature, and video-content recommendation results on MicroLens-100K.Source: Table 12 · p. 19
  • Text-only features yield performance similar to itemID, while short video titles may limit competitiveness and original videos contain much more information than titles.The authors suggest that stronger video-understanding techniques may improve recommendation results in the future.Source: J Comparison between Textual Features… · p. 19

K Recommendation Scenario of Collected Platform

Related papers

Loading 2309.15379v1…