Abstract
Graph Neural Networks (GNNs) often struggle with "information bottlenecks" by treating edges merely as structural links rather than rich data sources. We present the General Information Propagation Algorithm (GIPA), a novel architecture that elevates edge features to first-class citizens. By replacing standard attention mechanisms with a flexible Multi-Layer Perceptron (MLP) and integrating edge data directly into the propagation step, GIPA captures complex, non-linear relationships in dense graphs. Evaluated on the competitive Open Graph Benchmark (OGB) proteins dataset, GIPA achieved state-of-the-art results, outperforming existing models like UniMP and GAT.
Introduction
While GNNs have revolutionized structured data analysis, they frequently overlook the vital information contained within connections. In domains like protein-protein interaction networks, the data associated with an edge—the nature of the interaction—is often as critical as the node attributes themselves. Traditional attention mechanisms, such as those in Graph Attention Networks (GAT), typically rely on simple dot-products or single-layer projections, which may lack the expressivity needed for massive, dense graphs. GIPA addresses these limitations by providing a more general propagation framework that handles both node and edge features with high expressivity.
Methodology: The GIPA Process
The GIPA architecture is built on three innovative pillars designed to maximize information flow and relational modeling.
- MLP-Based Multi-Head Attention: Instead of simple dot-products, GIPA employs an MLP to learn complex, non-linear attention weights. The input is a concatenation of source node, target node, and edge features.
- Edge-Wise Propagation: GIPA’s propagation function combines a neighbor’s embedding with the specific edge embedding connecting it to the target node, ensuring the "message" carries the full context of the relationship.
- Residual Aggregation: To mitigate vanishing gradients in deep architectures (GIPA uses a 6-layer stack), the model uses residual connections. It concatenates aggregated messages with a linear projection of the target node’s original features before a final MLP update.
Experimental Setup
The model was evaluated using the ogbn-proteins dataset from the Open Graph Benchmark. This dataset is characterized by its significant scale and density, containing 132,534 nodes and over 79 million edges. Each edge carries 8-dimensional features representing biological interactions. The task is a multi-label binary classification involving 112 different protein functions, making it a rigorous test for edge-aware propagation.
Results and Discussion
GIPA achieved a new state-of-the-art performance on the ogbn-proteins leaderboard. The results highlight that in dense graphs, edge features contain a majority of the discriminative signal. By performing propagation "edge-wise," GIPA captures nuances missed by node-centric models.
As shown in the table above, GIPA reached a Test ROC-AUC of 0.8700 ± 0.0010, surpassing previous top-tier models. The use of concatenation over summation in the aggregation step proved effective in preserving node identity and preventing "over-smoothing," where node embeddings become indistinguishable in deep networks.
Conclusion and Future Work
GIPA demonstrates that treating edges as primary features significantly enhances GNN performance on complex relational data. While currently optimized for homogeneous graphs, future work will extend GIPA to heterogeneous networks with multiple entity and relation types. Additionally, the researchers aim to explore more efficient sampling methods to scale the MLP-based attention mechanism to even larger "web-scale" graphs.
Summary updated