GCN-LASE: Towards Adequately Incorporating Link Attributes in Graph Convolutional Networks

Ziyao Li, Liang Zhang, Guojie Song

Abstract

Standard Graph Convolutional Networks (GCNs) typically treat edges as binary indicators or scalar weights, overlooking the rich, multi-dimensional information often present in real-world connections. We present GCN-LASE, a framework that incorporates vector-valued link attributes into the learning process. By utilizing tensor products to model interactions between node and link features, LASE captures complex relational contexts. Our approach, derived from graph kernel theory, introduces a modular architecture consisting of Gate, Amplifier, and Aggregator components. Experimental results on social, citation, and temporal networks demonstrate that GCN-LASE significantly outperforms traditional GCNs and GraphSAGE, particularly in scenarios where link attributes provide critical context.

Introduction

In graph-based machine learning, the focus has traditionally been on node features. However, in many domains, the relationships themselves carry significant data. A link in a social network might represent a family tie, a professional collaboration, or a casual acquaintance—each with distinct implications for information flow. Most existing GCN architectures simplify these links into binary values or single weights, losing the nuance of the underlying data.

What Problem Does This Paper Address?

The primary challenge lies in how to mathematically integrate multi-dimensional link attributes. Simple methods like concatenating link features to node features often fail to capture how a specific relationship type modifies the neighbor's influence.

Examples of graphs with rich, vector-valued link attributes, such as shopping item similarities and temporal contact sequences.

As shown above, link attributes can represent diverse information, from content similarity in shopping networks to time-series data in temporal phone contact graphs. The authors seek a principled way to allow the network to "understand" the specific context of every connection.

Methods: The LASE Framework

The authors propose LASE (Link Attributes and Sampling Estimation), which introduces three key innovations to address the limitations of standard GCNs.

1. Neighbor Feature Tensors

Instead of additive or concatenative methods, LASE employs the tensor product of node and link features. This creates a "fully associated" feature representation that captures all bilinear combinations of attributes, ensuring that the interaction between the node's state and the link's type is explicitly modeled.

2. Kernel-Derived Architecture

The LASE architecture is grounded in established graph kernels, specifically Random Walk and Weisfeiler-Lehman kernels. This theoretical foundation leads to a modular design:

  • Gate: Evaluates the overall influence or importance of a neighbor.
  • Amplifier: Uses link information to element-wise modify neighbor node attributes, acting as a context-aware filter.
  • Aggregator: Combines these modified neighbor embeddings into the target node's representation.
The LASE architecture (top) and a structural comparison (bottom) demonstrating why simple feature concatenation is insufficient.

3. Min_Var Sampling

To maintain scalability on large graphs, LASE utilizes a "minimum variance" Monte Carlo sampling strategy. By calculating optimal sampling probabilities every $k$ batches, the model balances training efficiency with estimation accuracy, leading to more stable convergence than uniform sampling.

Key Insights

A critical finding is the failure of concatenation. As illustrated in the structural comparison above, models that simply append link features to node features cannot distinguish between different graph configurations where the same set of features is present but distributed differently across edges. The tensor product approach in LASE ensures the model can "see" the specific pairing of node and link attributes.

Furthermore, the authors demonstrate that temporal graphs can be effectively modeled by treating the time-series of connections as a vector attribute on a static edge, providing a powerful alternative to complex dynamic graph architectures.

Experimental Results

LASE was evaluated on four diverse datasets: Reddit (social), DBLP (citation), and Email/fmobile (temporal).

  • Superior Accuracy: LASE-SAGE and LASE-RW consistently outperformed standard GCNs. On the email dataset, LASE-SAGE achieved a Micro-f1 score of 0.7310, compared to 0.6548 for GraphSAGE.
  • Robustness to Noise: Signal-to-Noise Ratio analysis confirmed that the model's performance is directly tied to the quality of link data, proving it successfully leverages the provided attributes.

Conclusion

GCN-LASE represents a significant step toward more expressive graph neural networks. By treating links as rich context rather than simple pipes, the model achieves state-of-the-art performance on complex, attribute-heavy graphs. While computational complexity remains a challenge for industrial-scale deployment, the use of tensor products and principled sampling offers a robust framework for future research in relational learning.

Summary updated

Abstract

Graph Convolutional Networks (GCNs) have proved to be a most powerful architecture in aggregating local neighborhood information for individual graph nodes. Low-rank proximities and node features are successfully leveraged in existing GCNs, however, attributes that graph links may carry are commonly ignored, as almost all of these models simplify graph links into binary or scalar values describing node connectedness. In our paper instead, links are reverted to hypostatic relationships between entities with descriptional attributes. We propose GCN-LASE (GCN with Link Attributes and Sampling Estimation), a novel GCN model taking both node and link attributes as inputs. To adequately captures the interactions between link and node attributes, their tensor product is used as neighbor features, based on which we define several graph kernels and further develop according architectures for LASE. Besides, to accelerate the training process, the sum of features in entire neighborhoods are estimated through Monte Carlo method, with novel sampling strategies designed for LASE to minimize the estimation variance. Our experiments show that LASE outperforms strong baselines over various graph datasets, and further experiments corroborate the informativeness of link attributes and our model's ability of adequately leveraging them.
BibTeX
@inproceedings{li2019gcn,
  author = {Ziyao Li and Liang Zhang and Guojie Song},
  title = {{GCN-LASE: Towards Adequately Incorporating Link Attributes in Graph Convolutional Networks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/410},
}
Figure