Standard Graph Neural Networks (GNNs) rely heavily on the assumption of homophily—the principle that connected nodes tend to share the same class or attributes. Under this assumption, traditional neighborhood aggregation acts as a localized smoothing operator, averaging out differences to denoise node representations.
However, many real-world networks exhibit heterophily, where "opposites attract." In a fraud network, for instance, a fraudulent account may deliberately connect primarily to benign accounts to camouflage its behavior. When a standard GNN applies uniform neighborhood averaging in a heterophilic setting, it triggers a signal washing effect: the distinct features of minority classes are completely diluted by the overwhelming majority of surrounding connections.
To solve this, a new paradigm has emerged: label-wise neighborhood aggregation. Instead of treating a neighborhood as a single homogeneous mixture, these methods partition or filter neighbors based on their explicit, predicted, or soft class labels. By aggregating features within class-specific groups, these models preserve the local context and transform heterophily from a source of destructive noise into a highly discriminative structural signature.
Neighborhood Composition as a Semantic Signature
The core intuition behind label-wise aggregation is that heterophily is rarely random; it is often highly structured. Even if a node's neighbors belong to different classes than the node itself, nodes of the same category tend to share a consistent distribution of neighbor classes across the graph.
As identified by Jiahao Xu et al. in their work on HiGNN, this local neighborhood composition serves as a distinct semantic signature. For example, a "Class A" node might consistently connect to 70% "Class B" nodes and 30% "Class C" nodes. If a GNN simply averages these neighbors together, the unique identity of Class A is lost. But if the model maps out the specific mix of neighbor types, it can easily deduce the central node's true category.
By shifting the goal from smoothing features to profiling neighborhood distributions, models can successfully classify nodes even when immediate neighbors look entirely dissimilar.
Partitioning the Neighborhood by Class Labels
To capture these signatures without washing out critical signals, architectures must isolate the contributions of different neighbor classes. The most direct way to achieve this is through explicit neighborhood partitioning.
Standard Aggregation:
[Neighbor A (Class 1)] \
[Neighbor B (Class 2)] --> [ Average All ] --> Diluted Representation
[Neighbor C (Class 1)] /
Label-Wise Aggregation:
[Neighbor A (Class 1)] \ --> [ Aggregate Class 1 ] \
[Neighbor C (Class 1)] / --> [ Combine Groups ] --> Discriminative Representation
[Neighbor B (Class 2)] --> [ Aggregate Class 2 ] /
In LW-GCN, Enyan Dai et al. designed a mechanism that groups a node's neighbors according to their predicted class labels. The model performs separate feature aggregations for each label group before concatenating or combining the results. Because ground-truth labels are missing for unlabeled nodes during training, the authors utilize an initial predictive stage in HP-GMN to assign pseudo-labels, which then guide the class-wise feature grouping.
This partitioning strategy is particularly vital in highly imbalanced or adversarial domains like fraud detection. As Yuchen Wang et al. demonstrate with GAGA, splitting adjacent nodes into distinct groups based on their categories prevents minority signals (such as fraudulent activity) from being obscured by a massive majority of benign connections. By feeding these partitioned groups into a Transformer architecture, the model can intelligently weigh the relationships between different classes rather than blindly fusing them.
Guiding Information Flow with Soft Labels and Filters
Rather than relying on hard partitions, several approaches use continuous class probability distributions—soft labels—or adaptive node-specific filters to modulate how information propagates across class boundaries.
Instead of calculating graph attention solely from raw node features, Yubin Wang et al. developed SLGAT, which routes attention by combining node features with aggregated soft label distributions from the neighborhood. This ensures that the attention mechanism explicitly factorizes the underlying class structure, preventing the blurring of representations across distinct class boundaries.
Other methods explicitly vectorize these neighborhood distributions to treat them as primary features:
- Distribution Vectors: Zhuonan Zheng et al. introduced Heterophilous Distribution Propagation (HDP), which maps the statistical distribution of neighbor classes directly into a vector representation. This treats the variety of neighbor types as an explicit, discriminative feature set for classification.
- Node-Specific Filtering: In LACA, Shixiong Jing et al. employ node-specific filters that dynamically adapt feature aggregation based on neighbor label information, ensuring that majority class signals do not dominate the local representation.
- Multi-Hop Vicinity Analysis: Expanding beyond immediate neighbors, Shuichiro Haruta et al. developed EVAGNN, which analyzes the distribution of features across multiple neighborhood hops. This multi-hop context allows the model to differentiate between nodes that might appear identical under a shallow, single-hop aggregation scheme.
Why It Works: The Stability of Structured Heterophily
Label-wise neighborhood aggregation succeeds because it leverages underlying mathematical and structural regularities that traditional GNNs ignore.
First, neighborhood aggregation possesses an inherent concentration property. As proven by Haonan Wang et al. in their work on SP-GCL, neighborhood aggregation acts as a label-consistent filter that stabilizes node representations within the same class. Even when immediate neighbors have entirely different labels, the expected value of the aggregated features remains statistically consistent for nodes belonging to the same category.
Second, heterophily is beneficial when the connection preferences between classes are strong and predictable. By analyzing a wide array of heterophilous GNNs under a unified framework, Zhuonan Zheng et al. demonstrated in their study of CMGNN that these models succeed by explicitly or implicitly enhancing the compatibility matrix between node classes.
When a model isolates neighbor features by label, it learns specialized filters tailored to each cross-class interaction (e.g., how a Class A node interacts specifically with a Class B neighbor). This allows the network to exploit predictable patterns of dissimilarity, turning what standard message-passing perceives as chaotic noise into a highly structured, predictive asset.
Key Foundational Work
Label-Wise Graph Convolutional Network for Heterophilic Graphs (2022) — contributed 3 concept elements to this direction.
HP-GMN: Graph Memory Networks for Heterophilous Graphs (2022) — contributed 3 concept elements to this direction.
SLGAT: Soft Labels Guided Graph Attention Networks (2020) — contributed 3 concept elements to this direction.