Skip to main content
Back

NYU Abu Dhabi Pits Quantum-Inspired Node Embeddings Against Classical Graph AI

A controlled benchmark pits classical node embeddings against quantum-oriented ones inside a fixed GNN pipeline, revealing dataset-dependent winners rather than a clear champion.

Paper, Full Access
  • Controlled benchmark compares classical and quantum-oriented node embeddings in a fixed GIN pipeline (paper).
  • Six embeddings tested: Fixed, MLP, Angle-VQC, QuOp, QWalkVec, QPE across five TU datasets plus QM9.
  • QWalkVec* wins MUTAG (0.947), PROTEINS (0.777), QM9 (0.852) when paired with a learnable projection.
  • Classical MLP still wins attribute-scarce IMDB-BINARY; Angle-VQC collapses there to 0.60 accuracy.
  • Trainability is non-uniform: essential for QWalkVec, neutral or harmful for QuOp depending on dataset.
  • No code release; accepted at IJCNN 2026 with DOI 10.1109/IJCNN69854.2026.11728756.

A controlled graph-learning study examines how node embeddings affect graph-level predictions when the backbone, data splits, optimizer, and training budget remain fixed. Researchers at NYU Abu Dhabi and Sapienza University of Rome find that rankings depend heavily on the dataset: walk-based embeddings lead several molecular and protein tasks, while simpler alternatives remain competitive or win on social graphs.

The paper, accepted at IJCNN, compares two classical baselines, one variational quantum circuit, and three quantum-inspired methods inside the same Graph Isomorphism Network, or GIN. This design gives developers a clearer view of whether performance changes come from the embedding module rather than a different model or training recipe.

One backbone, fewer confounders

A node embedding converts each node’s raw attributes and structural information into the vector consumed by a GNN’s message-passing layers. Changing that vector can alter what the network learns about local neighborhoods, graph position, and multi-hop structure.

Earlier comparisons often changed several components at once, including the GNN architecture, split strategy, optimizer settings, and epoch budget. The new benchmark holds those choices constant across methods. Random variation and dataset-specific effects still remain, but the controls make embedding-level comparisons more direct.

Six encoders, one input recipe

Unified evaluation pipeline for six node-embedding methods
Each embedding module feeds the same GIN backbone and evaluation pipeline.

Every node begins with a descriptor that concatenates a one-hot degree encoding with Laplacian positional encodings, which use the graph spectrum to represent structural position. The six methods transform that shared input in different ways:

  • Fixed: applies a non-trainable random projection.
  • MLP: learns a feed-forward transformation of the base features.
  • Angle-VQC: encodes features as rotation angles in a variational quantum circuit, applies entangling layers, and reads Pauli-Z expectation values.