Jumping Knowledge Networks
What you'll learn: How Jumping Knowledge Networks let each node choose its own ideal receptive field size by combining information from multiple GNN layers.
The Problem with Fixed-Depth GNNs
In standard Graph Neural Networks, every node uses the same number of message-passing layers. This means all nodes have the same receptive field size — they all look at neighborhoods of the same "distance." But in real graphs, some nodes benefit from local information (1–2 hops away), while others need broader context (3–4 hops). A fixed depth forces a one-size-fits-all solution.
The Jumping Knowledge Idea
Jumping Knowledge Networks (JK-Nets) solve this by giving each node access to representations from all intermediate layers, then letting the node adaptively select or combine the most useful scale of information.
Think of it like adjusting a camera zoom: some subjects look best up close, others from far away. JK-Nets let each node "choose its zoom level."
How It Works
After running multiple GNN layers (e.g., 1, 2, 3 hops), instead of using only the final layer's output, JK-Nets collect representations from every layer. Then they apply a layer aggregation strategy:
- Concatenation: Stack all layer outputs together
- Max-pooling: Take element-wise maximum across layers
- LSTM-attention: Use a learned mechanism to weight different layers
This creates skip connections similar to ResNet, but specifically designed to capture multi-scale neighborhood information.
Why It Matters
Different node types in the same graph often need different receptive fields. Social network hubs benefit from wide views; isolated specialists need local focus. JK-Nets adapt automatically during training, improving performance on heterogeneous graph structures.
Key Takeaway: Jumping Knowledge Networks use skip connections and layer aggregation to let each node adaptively select the best receptive field size from multiple GNN layers, rather than forcing all nodes to use the same neighborhood scale.