Abstract
This paper proposes a Metapath-Infused Exponential Decay graph neural network (MIED) approach for node embedding in heterogeneous graphs. It is designed to address limitations in existing methods, which usually lose the graph information during feature alignment and ignore the different importance of nodes during metapath aggregation. Firstly, graph convolutional network (GCN) is applied on the subgraphs, which is derived from the original graph with given metapaths to transform node features. Secondly, an exponential decay encoder (EDE) is designed, in which the influence of nodes on starting point decays exponentially with a fixed parameter as they move farther away from it. Thirdly, a set of experiments is conducted on two selected datasets of heterogeneous graphs, i.e., IMDb and DBLP, for comparison purposes. The results show that MIED outperforms selected approaches, e.g., GAT, HAN, MAGNN, etc. Thus, our approach is proven to be able to take full advantage of graph information considering node weights based on distance aspects. Finally, relevant parameters are analyzed and the recommended hyperparameter setting is given.
Introduction
Heterogeneous graphs (HGs), which contain multiple types of nodes and edges, are essential for representing complex real-world systems like academic networks or movie databases. To learn effective node embeddings in these graphs, researchers often rely on metapaths—predefined sequences of node types (e.g., Author-Paper-Venue). However, current state-of-the-art models face two significant challenges:
- Information Loss during Alignment: Different node types often reside in different feature spaces. Standard models typically use simple linear projections to align these features, ignoring the local graph structure during this critical initial step.
- The Distance Problem: When aggregating information along a metapath, many models treat all nodes as equally important or use computationally heavy Recurrent Neural Networks (RNNs) that struggle with long-range dependencies.
The Metapath-Infused Exponential Decay (MIED) model addresses these hurdles by integrating graph-aware feature alignment and a novel distance-based encoding mechanism.
The MIED Framework
The MIED architecture consists of three primary stages designed to preserve structural context and model the relative importance of nodes based on their proximity.
GCN-based Node Feature Alignment
Instead of a simple linear transformation, MIED applies a Graph Convolutional Network (GCN) to subgraphs derived from the specified metapaths. This ensures that when features are projected into a unified embedding space, they already incorporate information from their immediate neighborhood, preserving the local topology that linear layers often discard.
Exponential Decay Encoder (EDE)
The core innovation of MIED is the Exponential Decay Encoder. It operates on the intuition that a node's influence on the target node should decrease as the distance between them in a metapath increases. By using a decay parameter ($\alpha$), the model exponentially reduces the weight of distant nodes. This approach is significantly more memory-efficient than RNN-based encoders, reducing path encoding storage from $O(K \cdot n)$ to $O(1)$. Experimental analysis suggests an optimal range for $\alpha$ between $1/2$ and $2/3$.
Metapaths Aggregation
Finally, MIED employs a hierarchical attention mechanism to aggregate information across different metapaths. This allows the model to dynamically learn which metapaths are most relevant for a specific node or task (e.g., prioritizing "Movie-Director-Movie" over "Movie-Genre-Movie" for genre classification).
Experimental Results
The model was evaluated on two benchmark datasets: IMDb (Movies, Directors, Actors) and DBLP (Authors, Papers, Terms, Venues).
Performance Comparison
MIED consistently outperformed strong baselines, including GAT, HAN, and MAGNN, across both node classification and clustering tasks:
- Node Clustering: On IMDb, MIED achieved a Normalized Mutual Information (NMI) of 15.84%, compared to MAGNN’s 15.58%. On DBLP, it reached an NMI of 81.84%.
- Node Classification: In Macro-F1 scores, MIED showed superior accuracy. For instance, on IMDb with a 20% training split, MIED scored 60.52%, nearly a 2% improvement over the previous state-of-the-art.
Ablation Study
Ablation experiments confirmed that both the GCN-based alignment and the EDE are critical components. Removing either led to a measurable decline in performance, proving that structural awareness and distance-based weighting are both necessary for high-quality embeddings.
Conclusion and Future Work
MIED provides a lean, efficient, and structure-aware framework for heterogeneous graph embedding. By replacing complex RNNs with a simple exponential decay heuristic and making feature alignment graph-aware, it achieves state-of-the-art results with lower computational overhead. Future research will focus on automating metapath selection and scaling the model for massive, web-scale recommendation systems.
Summary updated