A Vlogger-augmented Graph Neural Network Model for Micro-video Recommendation

Weijiang Lai, Beihong Jin, Beibei Li, Yiyuan Zheng, Rui ZhaoLecture Notes in Computer Science 2023 · DOI · arXiv · OpenAlex

Abstract

Micro-video platforms like TikTok and WeChat Channels rely heavily on content creators, or vloggers, to drive user engagement. However, traditional recommendation models often overlook the vlogger's role, focusing primarily on user-video interactions. This paper introduces VA-GNN (Vlogger-Augmented Graph Neural Network), a novel framework that models the tripartite relationship between users, videos, and vloggers. By leveraging dual-view embedding propagation and cross-view contrastive learning, VA-GNN captures both content-based and creator-centric preferences. Experimental results on large-scale datasets demonstrate significant improvements over state-of-the-art baselines, highlighting the importance of vloggers as first-class citizens in recommendation systems.

Introduction

In the ecosystem of micro-video platforms, vloggers are the primary content producers and often define the "soul" of the platform. Users frequently follow specific creators because they trust their style or niche (e.g., fitness, cooking, tech). Despite this, existing recommendation systems typically treat vlogger information as a static attribute or ignore it entirely, focusing instead on the bipartite interaction between users and individual videos.

This paper addresses the challenge of integrating complex, multi-layered relationships—user-video, user-vlogger, and vlogger-video—into a unified recommendation model. The authors argue that a user’s decision to engage with a video is often driven by their affinity for the creator, necessitating a model that can adaptively balance content signals with creator-centric signals.

Methodology: The VA-GNN Framework

The VA-GNN architecture, illustrating the tripartite graph construction, dual-view propagation, and contrastive learning components.

The proposed VA-GNN model consists of four primary stages designed to capture the nuances of vlogger-driven recommendations:

  1. Tripartite Graph Construction: The model builds a heterogeneous graph with three node types: Users ($u$), Micro-videos ($v$), and Vloggers ($p$). Edges represent interactions such as "likes," "follows," and "publishing" relationships.
  2. Dual-View Embedding Propagation: To capture different facets of user preference, the model learns representations from two perspectives:
  3. Video-view: Focuses on direct user-video interactions.
  4. Vlogger-view: Focuses on creator affinity using "meta-paths" (e.g., $user \to vlogger \to user$) generated via random walks. This allows the model to identify users with similar creator-based interests.
  5. Cross-view Contrastive Learning: To ensure consistency between the two views, the model employs contrastive learning. This aligns the latent features of a user across both the video and vlogger views, ensuring that "who you follow" and "what you watch" provide a coherent representation of user intent.
  6. Adaptive Gating Mechanism: During the prediction phase, the model uses a learned gate to adaptively weight the importance of the video-view versus the vlogger-view. This allows the system to prioritize creator affinity for loyal followers while relying on content signals for discovery.

Experimental Evaluation

Datasets

The model was evaluated on two large-scale, real-world datasets:

  • WeChat-Channels: A public dataset from the 2021 WeChat Big Data Challenge, containing 14 days of interaction data.
  • TakaTak: An industrial dataset from a popular Indian micro-video platform, covering four weeks of user behavior.

Key Results

VA-GNN consistently outperformed state-of-the-art GNN models (e.g., LightGCN) and contrastive learning methods (e.g., SimGCL).

  • Performance Gains: On the TakaTak dataset, VA-GNN achieved a 25.15% improvement in Recall@10 and a 32.78% improvement in NDCG@10 over the strongest baseline.
  • Ablation Studies: Removing the vlogger-view propagation or the vlogger-specific loss functions led to a significant drop in performance, confirming that modeling creator affinity is the primary driver of the model's success.
  • Robustness: Unlike models that rely on random graph perturbations for contrastive learning, VA-GNN’s use of structural views (video vs. vlogger) provides more stable and interpretable embeddings.

Conclusion and Future Work

VA-GNN demonstrates that vloggers act as essential "anchors" of consistency in micro-video recommendation. By treating them as active nodes in a tripartite graph and aligning user preferences through contrastive learning, the model achieves superior accuracy.

Future research directions include addressing the "cold start" problem for new vloggers with no initial followers and integrating multi-modal features—such as video thumbnails, audio, and vlogger descriptions—directly into the graph neural network to further refine node embeddings.

Summary updated

BibTeX
@inproceedings{lai2023vlogger,
  author = {Weijiang Lai and Beihong Jin and Beibei Li and Yiyuan Zheng and Rui Zhao},
  title = {{A Vlogger-augmented Graph Neural Network Model for Micro-video Recommendation}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2023},
  doi = {10.1007/978-3-031-43427-3\_41},
}
Figure