Jumping Knowledge network
Architectures AIMech:JumpingKnowledgeNetwork DRAFT
Aliases: JK-Net, Jumping Knowledge networks
A graph neural architecture that sends intermediate node representations from multiple neighborhood-aggregation depths to a final layer-aggregation mechanism.
Identity Notes & Concept Scope
Jumping Knowledge adds aggregation across layer depths and can be combined with GCN, GraphSAGE or GAT neighborhood operators. It is distinct from ordinary adjacent-layer skip connections and from any one of those base operators.
Claims & Grounded Evidence
1 Claim
Mechanism
#architecture-mechanism
Retain each node's intermediate neighborhood-aggregation representations and combine them at the final stage using concatenation, element-wise maximum, or an LSTM-based attention mechanism over depths.
⚠️ Curated Limitations & Negative Results
- Concatenation followed by a node-shared linear transformation is not node-adaptive in its layer weighting; max pooling and the attention variant provide node-dependent selection or weights.
- The influence-distribution result is conditional on activation assumptions. No general guarantee of optimal neighborhood range or benchmark superiority is asserted.
- The record describes the layer-aggregation family. The paper's attention prose refers to both weighted layer features and weighted bidirectional-LSTM features, so an exact attention implementation would require a separately pinned implementation review.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
broad_match | The source builds a neural architecture from neighborhood-aggregation layers and a final learned or pooling combination; the artificial-neural-network target is broader. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, Stefanie Jegelka
|
primary paper
full_text
|
sha256:4aa02e943446…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:11:44Z
Updated: 2026-10-11T04:11:44Z
Source File:
data/curated/JumpingKnowledgeNetwork.yaml