A2-GCN: An attribute-aware attentive GCN model for recommendation

Fan Liu, Zhiyong Cheng, Lei Zhu, Chenghao Liu, Liqiang NieOpenReview 2022 · OpenReview

TL;DR

This paper introduces A2-GCN, a novel recommendation model that leverages item attributes more effectively, especially when attributes are missing. By constructing a graph of users, items, and attributes and employing an attention mechanism, A2-GCN significantly improves recommendation accuracy and robustly handles incomplete attribute data.

What Problem Does This Paper Address?

Recommender systems often struggle with sparse data, meaning users have interacted with only a small fraction of available items. To combat this, researchers have incorporated "side information" like item attributes (e.g., movie genres, product brands). However, a major challenge is that these attributes are frequently missing in real-world datasets. Existing methods often fill missing attributes with a generic placeholder like "Other" or simply ignore them, both of which can inject noise and lead to suboptimal recommendations. Furthermore, these methods typically treat all attributes equally, failing to recognize that different users might value different attributes of the same item.

Methods

The core of A2-GCN is a Graph Convolutional Network (GCN) built upon a heterogeneous graph. This graph connects three types of nodes: users, items, and attributes.

  • Graph Construction: Users are linked to items they've interacted with, and items are linked to their associated attributes.
  • Message Passing: The model uses a message-passing strategy to learn node representations. Each node aggregates information from its neighbors. For instance, an item's representation is learned by combining messages from its user and attribute neighbors.
  • Attribute-Aware Attention: A key innovation is an attribute-aware attention mechanism. This mechanism allows the model to dynamically weigh the importance of different attributes when learning user representations. It considers how an item's attributes influence a user's preference for that specific item. This is crucial because different users might focus on different aspects of an item (e.g., one user might care about a movie's director, another about its actors).
  • Handling Missing Attributes: By only using available attribute information and propagating it through the graph, A2-GCN naturally avoids making assumptions about missing attributes, thus preventing the introduction of misleading information.

Datasets

The authors evaluated their model on four publicly available Amazon product review datasets:

  • Office Products
  • Clothing
  • Toys Games
  • Kindle Store

These datasets vary in size, sparsity, and the number of attributes, allowing for a comprehensive evaluation across different scenarios.

Key Insights and Ideas

  • Attributes are powerful, but missing attributes are a major hurdle: The paper strongly emphasizes that while attributes significantly boost recommendation performance, the way existing methods handle missing attributes is a critical bottleneck.
  • Graph structure is key for attribute integration: Representing users, items, and attributes as a unified graph allows for a more holistic understanding of their interdependencies.
  • Attention is not one-size-fits-all: The proposed attribute-aware attention mechanism is crucial for capturing nuanced user preferences, recognizing that attribute importance is context-dependent (user-item specific).
  • GCNs are a natural fit for attribute-rich recommendation: The graph-based approach of GCNs is well-suited for propagating attribute information and learning rich representations.

Key Results

  • State-of-the-art performance: A2-GCN consistently outperformed several strong baselines (including those that also use attributes) across all four datasets, achieving significant improvements in Hit Ratio (HR@20) and Normalized Discounted Cumulative Gain (NDCG@20). For example, it achieved a relative improvement of up to 16.19% in HR@20 on the Clothing dataset compared to the best baseline.
  • Robustness to missing attributes: The model demonstrated remarkable resilience to missing attribute data. While other attribute-aware methods saw their performance degrade significantly as more attributes were removed, A2-GCN's performance remained much more stable, highlighting its effectiveness in handling incomplete information.
  • Effectiveness of attention: Ablation studies showed that both the general attention mechanism and the specific attribute-aware attention mechanism contributed positively to the model's performance, with the attribute-aware version yielding further gains.

Limitations and Future Work

  • Attribute embedding: The paper notes that while the model learns user and item embeddings, it does not explicitly learn attribute embeddings. The authors suggest that attribute embeddings could also be updated similarly to user and item embeddings, which could further enhance the model.
  • User attributes: The current model focuses on item attributes. The authors mention that the framework can be extended to incorporate user attributes (e.g., age, gender) for even richer representations.
  • Generalization: While the model performed well across diverse datasets, further exploration on datasets with different types of side information or more complex attribute relationships could be beneficial.

Summary updated

Abstract

As important side information, attributes have been widely exploited in the existing recommender system for better performance. In the real-world scenarios, it is common that some attributes of items/users are missing (e.g., some movies miss the genre data). Prior studies usually use a default value (i.e., “other”) to represent the missing attribute, resulting in sub-optimal performance. To address this problem, in this paper, we present an attribute-aware attentive graph convolution network (A2-GCN). In particular, we first construct a graph, whereby users, items, and attributes are three types of nodes and their associations are edges. Thereafter, we leverage the graph convolution network to characterize the complicated interactions among <users, items, attributes>. To learn the node representation, we turn to the message-passing strategy to aggregate the message passed from the other directly inked types of nodes (e.g., a user or an attribute). To this end, we are capable of incorporating associate attributes to strengthen the user and item representations, and thus naturally solve the attribute missing problem. Considering the fact that for different users, the attributes of an item have different influences on their preference for this item, we design a novel attention mechanism to filter the message passed from an item to a target user by considering the attribute information. Extensive experiments have been conducted on several publicly accessible datasets to justify our model. Results show that our model outperforms several state-of-the-art methods and demonstrates the effectiveness of our attention method.
BibTeX
@misc{liu2022gcn,
  author = {Fan Liu and Zhiyong Cheng and Lei Zhu and Chenghao Liu and Liqiang Nie},
  title = {{A2-GCN: An attribute-aware attentive GCN model for recommendation}},
  year = {2022},
  url = {https://openreview.net/forum?id=QuDfBKjY3tf},
  howpublished = {OpenReview},
}
Figure