Multi-View Graph Convolutional Network for Multimedia Recommendation

Penghang Yu, Zhiyi Tan, Guanming Lu, Bing-Kun BaoACM Multimedia 2023

TL;DR

This paper introduces MGCN, a novel Multi-View Graph Convolutional Network designed to improve multimedia recommendation systems. MGCN tackles the issues of modality noise contamination and incomplete user preference modeling by purifying modality features with behavior information and adaptively fusing them based on user preferences, leading to significant performance gains.

What Problem Does This Paper Address?

Recommender systems are ubiquitous, aiming to help users discover relevant items. In recent years, multimedia recommendation has gained traction by incorporating item multimodal information (like images and text) alongside user behavior data. While Graph Convolutional Networks (GCNs) have shown promise in this area, existing methods face two major challenges:

  1. Modality Noise Contamination: Raw multimodal features often contain irrelevant information (e.g., background noise in images, redundant text). When these noisy features are directly fused with behavior features in GCNs, the noise can be amplified and coupled with genuine preferences, leading to less discriminative item representations and poorer recommendation quality.
  2. Incomplete User Preference Modeling: Current approaches tend to fuse modality features equally, ignoring the fact that users have varying preferences for different modalities depending on the item. For instance, a user might prioritize visual appeal for clothing but textual descriptions for electronics. This "one-size-fits-all" fusion leads to suboptimal user preference modeling.

Methods

The proposed Multi-View Graph Convolutional Network (MGCN) addresses these challenges with three key modules:

  • Behavior-Guided Purifier: This module aims to mitigate modality noise contamination. It uses item behavior information to filter out preference-irrelevant features from raw modality features, ensuring that only the most relevant information is retained.
  • Multi-View Information Encoder: This module enhances feature discriminability by capturing two types of signals in separate "views":
  • User-Item View: A GCN is used to capture high-order collaborative signals from user-item interactions.
  • Item-Item View: Another GCN operates on item-item affinity graphs (constructed using cosine similarity of raw modality features and sparsified using KNN) to capture semantically correlative signals between items.
  • Behavior-Aware Fuser: This module adaptively fuses the purified and enriched modality features with behavior features. It distills user modality preferences from their behavior history and uses these preferences to learn adaptive fusion weights. Additionally, a self-supervised auxiliary task is introduced to maximize the mutual information between fused multimodal features and behavior features, encouraging the model to capture both complementary and supplementary preference information.

Datasets

The authors evaluated their model on three publicly available Amazon datasets:

  • Baby
  • Sports and Outdoors
  • Clothing, Shoes, and Jewelry

These datasets are commonly used in multimedia recommendation research. The paper notes that pre-extracted visual (4096-dimensional) and textual (384-dimensional) features were used.

Key Insights and Ideas

  • Noise is a significant problem: Directly injecting raw multimodal features into GCNs for recommendation is detrimental due to inherent noise, leading to performance degradation.
  • Behavior is key to purification: User behavior data is crucial not just for modeling preferences but also for guiding the purification of noisy modality features.
  • Multi-view learning is beneficial: Separately capturing collaborative (user-item) and semantic (item-item) signals in different graph views enhances feature discriminability.
  • User preferences are heterogeneous: Users don't treat all modalities equally. An adaptive fusion mechanism that considers individual user preferences for different modalities is essential for effective recommendation.
  • Self-supervised learning aids fusion: An auxiliary self-supervised task can further improve the fusion process by encouraging alignment between multimodal and behavior representations.

Key Results

  • Overall Superiority: MGCN significantly outperforms both general recommendation models (like MF, LightGCN) and existing multimedia recommendation models (VBPR, MMGCN, GRCN, SLMRec, BM3, MICRO) across all three datasets.
  • Significant Improvement: On the Clothing dataset, MGCN achieved a 23.3% improvement over the best baseline methods.
  • Module Importance: Ablation studies showed that removing any of the core modules (Behavior-Guided Purifier, Multi-View Information Encoder, Behavior-Aware Fuser) led to a noticeable drop in performance, highlighting their individual contributions.
  • Modality Impact: Visual features were found to have a greater impact on recommendation performance than textual features, likely because users often make purchasing decisions based on appearance.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that GCN-based methods can be prone to over-smoothing, although they mitigate this by using a shallow GCN in the item-item view.
  • Some methods like GRCN and MICRO require frequent graph structure updates, leading to high computational complexity. MGCN's purifier approach avoids this.
  • Future Work: The authors suggest integrating external knowledge with item information through large-scale language models to address the cold-start problem.

Summary updated

BibTeX
@inproceedings{yu2023multi,
  author = {Penghang Yu and Zhiyi Tan and Guanming Lu and Bing-Kun Bao},
  title = {{Multi-View Graph Convolutional Network for Multimedia Recommendation}},
  booktitle = {ACM Multimedia},
  year = {2023},
}
Figure