Standard Graph Neural Networks (GNNs) typically employ a fixed number of message-passing layers, applying a uniform "receptive field" to every node in a graph. However, graph structures are inherently non-uniform. A hub node in a dense cluster requires a different level of context than a peripheral node in a sparse region. Applying the same depth to both leads to two primary failures: over-smoothing, where node features become indistinguishable in dense regions, and under-reaching, where nodes in sparse regions fail to capture necessary long-range dependencies.
To address this, researchers have developed mechanisms for adaptive aggregation depth, allowing the model to dynamically determine how many "hops" of information each node or edge requires.
Multi-Scale Feature Integration
The earliest approaches to varying neighborhood influence focused on integrating representations from all intermediate layers rather than just the final one. Keyulu Xu et al. introduced Jumping Knowledge (JK) Networks in Representation Learning on Graphs with Jumping Knowledge Networks, which allow each node to "jump" its representations from different depths to a final aggregation layer.
By using a max-pooling mechanism across these layers, the model adaptively selects the most significant features from various neighborhood ranges for each individual node. This decouples the depth of the network from the range of neighborhood influence, preventing information loss in dense regions where global aggregation might otherwise become redundant and noisy.
Dynamic Halting and Reinforcement Learning
A more direct way to achieve adaptive depth is to allow nodes to stop aggregating information once they have reached a stable representation. Indro Spinelli et al. proposed the Adaptive Propagation Graph Convolutional Network (AP-GCN) in Adaptive Propagation Graph Convolutional Network, which incorporates a differentiable "halting unit." This unit outputs a stopping probability at each iteration, enabling the model to learn a per-node propagation depth that balances communication depth against the risk of over-smoothing.
Other researchers have framed this as a sequential decision problem. Daochen Zha et al. developed Policy-GNN in Policy-GNN: Aggregation Optimization for Graph Neural Networks, using deep reinforcement learning to determine the optimal number of hops for each node. Similarly, Xusheng Zhao et al. applied this to brain network analysis in Deep Reinforcement Learning Guided Graph Neural Networks for Brain Network Analysis, showing that dynamic depth allows models to capture long-range interactions for complex graph instances while avoiding noise in simpler ones.
Navigating Heterophily
Adaptive depth is particularly critical in heterophilic graphs, where connected nodes often belong to different classes (e.g., in fraud detection or protein interactions). In these settings, deep aggregation often introduces "noise" by washing out a node's unique features with those of its dissimilar neighbors.
Asela Hevapathige argues in Beyond Fixed Depth: Adaptive Graph Neural Networks for Node Classification Under Varying Homophily that node-wise adaptive depth allows a model to adapt to varying levels of homophily across different regions of the same graph. By restricting aggregation depth in heterophilic regions, the model preserves local signals, while allowing deeper exploration in homophilic regions to maximize the utility of reliable structural information. Xin Zheng et al. further automated this in Auto-HeG: Automated Graph Neural Network on Heterophilic Graphs, using neural architecture search to discover layer-specific strategies that prioritize different neighborhood ranges.
Continuous and Stochastic Depth
Recent work has moved beyond discrete layer counts to treat depth as a continuous or probabilistic parameter. Huiyuan Chen et al. proposed TEDGCN in From Trainable Negative Depth to Edge Heterophily in Graphs, which parameterizes depth as a continuous real number. This allows the model to optimize the degree of signal propagation via backpropagation, effectively switching between low-pass (smoothing) and high-pass (sharpening) filtering.
Alternatively, Paribesh Regmi treats neighborhood scope as a stochastic process in Bayesian Neighborhood Adaptation for Graph Neural Networks. By assigning a contribution probability to each hop, the network automatically infers the optimal neighborhood scope during training, eliminating the need for manual hyperparameter tuning of layer counts.
Efficiency and Scalability
Beyond accuracy, adaptive depth offers significant computational advantages. Wentao Zhang et al. demonstrated in Accelerating Scalable Graph Neural Network Inference with Node-Adaptive Propagation that customizing propagation depth during inference can lead to up to a 75× speedup. By terminating computations for nodes that reach stable representations at shallower depths, the model avoids redundant calculations on massive, web-scale graphs without sacrificing performance.
Key Foundational Work
Representation Learning on Graphs with Jumping Knowledge Networks (2018) — contributed 3 concept elements to this direction.