TL;DR
This paper introduces LatGRL, a novel unsupervised representation learning framework designed specifically to handle semantic heterophily in heterogeneous graphs. By constructing fine-grained homophilic and heterophilic latent graphs alongside an adaptive dual-frequency semantic fusion mechanism, LatGRL successfully captures both shared and distinct node characteristics without relying on labeled data.
What Problem Does This Paper Address?
Graph Neural Networks (GNNs) typically operate under the assumption of homophily—the idea that connected nodes share similar attributes or labels ("birds of a feather flock together"). However, real-world heterogeneous graphs (HGs), which contain multiple types of nodes and edges, frequently exhibit semantic heterophily. For example, in a movie database, a single actor might star in completely different genres of movies (e.g., action and comedy), meaning nodes connected via a "Movie-Actor-Movie" meta-path will have different labels.
Most existing Unsupervised Heterogeneous Graph Representation Learning (UHGRL) methods rely heavily on low-pass filtering, which smooths features across neighboring nodes. In heterophilic scenarios, this standard approach backfires by blurring the boundaries between different categories, severely degrading representation quality.
Methods
The authors propose LatGRL (Latent Graphs Guided Unsupervised Representation Learning) to tackle this challenge through three main modules:
- Latent Graphs Construction: Instead of relying purely on the original, potentially noisy graph topology, LatGRL mines node similarities by coupling global structure (using random walk probability diffusion across meta-paths) with node feature similarity. This allows the framework to construct a duo of latent graphs: a homophilic latent graph (connecting highly similar nodes) and a heterophilic latent graph (connecting structurally distant, dissimilar nodes).
- Adaptive Dual-Frequency Semantic Fusion: To prevent feature blurring, LatGRL applies both low-pass graph filtering (to capture commonalities within neighborhoods) and high-pass graph filtering (to sharpen differences across heterophilic edges). A node-wise attention mechanism adaptively fuses these dual-frequency representations based on each node's unique local neighborhood pattern.
- Latent Graphs Guided Learning: The framework maximizes the mutual information between the fused node representations and the latent graphs using an InfoNCE contrastive loss, effectively using the latent graphs as self-supervised category guides.
To scale up to massive real-world networks, the authors also developed LatGRL-S, a variant that optimizes latent graph construction using first-order neighbors and anchor nodes, combined with a decoupled feature pre-filtering step to achieve linear time complexity.
Datasets
The framework was evaluated on four standard benchmark datasets and one large-scale dataset:
- ACM: Academic paper network (Target: Papers, 3 classes)
- DBLP: Computer science bibliography network (Target: Authors, 4 classes)
- IMDB: Movie database network (Target: Movies, 3 classes)
- Yelp: Merchant review network (Target: Businesses, 3 classes)
- Ogbn-mag: A massive subset of the Microsoft Academic Graph used to test scalability (Target: Papers, 349 classes)
Key Insights and Ideas
- Neighborhood Diversity: The authors show that homophily is not just a graph-level or meta-path-level metric; it varies drastically at the individual node level. A single node can experience high homophily under one meta-path but high heterophily under another.
- High-Pass Filtering is Essential: While low-pass filtering is the default for most GNNs, high-pass filtering is crucial in heterophilic settings because it acts as a differentiator, sharpening the contrast between distinct categories in local neighborhoods.
- Structure + Features > Either Alone: Relying solely on graph structure or solely on node features for contrastive learning yields suboptimal positive samples. Coupling them via joint similarity mining provides a much cleaner self-supervised signal.
Key Results
- Downstream Performance: LatGRL consistently outperformed 11 state-of-the-art baselines (including supervised methods like HAN and GAT) in node classification and clustering tasks. The performance gains were particularly pronounced on highly heterophilic datasets like Yelp and IMDB.
- Qualitative Separation: t-SNE visualizations on the ACM dataset showed that LatGRL creates significantly tighter clusters with clearer, more distinct boundaries compared to competing methods like HeCo and DMGI.
- Linear Scalability: On the massive
ogbn-magdataset, LatGRL-S achieved superior accuracy while maintaining a fraction of the training time required by other unsupervised methods.
Limitations and Future Work
The authors note that the quality of the constructed latent graphs is inherently bounded by the quality of the initial node features. In datasets where the original features possess exceptionally weak discriminative capability (such as IMDB), the homophily ratio of the latent graphs can be negatively affected.
Potential future directions include exploring more sophisticated graph filtering techniques or extending the latent graph guidance mechanism to dynamic heterogeneous graphs.
Summary updated