Learning from Pure Topology in Unattributed Graphs

This direction is merged into another direction: attribute-agnostic graph representation. Its last version stays readable and citable here.

Version 1 of 182 papers96 conceptsfrozen 2026-10-04

In the standard Graph Neural Network (GNN) paradigm, we assume that nodes come with rich feature vectors—text in a citation network, chemical properties in a molecule, or user profiles in a social network. However, real-world data is often "content-less." In domains like botnet detection, biological interaction mapping, or sparse matrix partitioning, we may have only the raw connectivity: who is talking to whom, but not what they are saying.

Research into unattributed graph learning addresses this "featureless" reality. The core intuition is that the arrangement of connections—the topology—is itself a dense signal that can substitute for, or even outperform, external metadata.

Forcing a Structural Focus

When node attributes are missing, the first challenge is simply giving the GNN something to "message pass." A common but naive approach is to use random noise or mean values, but this provides no signal.

To isolate the influence of network connectivity, researchers like Vikash Singh and Pietro Lio in Towards Probabilistic Generative Models Harnessing Graph Neural Networks for Disease-Gene Prediction suggest using an identity matrix as node features. This ensures that the resulting embeddings are derived solely from the graph's structural topology. Similarly, Alexander M Rush and Jiawei Zhou demonstrate in their work on botnet detection that initializing node features as uniform constant vectors forces the model to recognize "fast-mixing" and hierarchical communication patterns rather than individual node identities.

The most extreme version of this intuition is found in the Graph Positional and Structural Encoder (GPSE). Renming Liu, Dominique Beaini, and colleagues pre-train an encoder on random Gaussian noise instead of node attributes. By forcing the model to reconstruct hand-crafted structural encodings from noise, they create a universal representation that transfers across vastly different domains—from chemistry to computer vision—proving that connectivity alone contains enough information to build a "foundational" graph model.

Synthetic Features as Structural Proxies

If raw topology is the signal, how do we represent it? One effective strategy is to "materialize" the structure into synthetic features.

Chi Thang Duong and Karl Aberer show in On Node Features for Graph Neural Networks that structural feature initialization—using metrics like degree centrality, PageRank, or DeepWalk embeddings—can serve as a powerful substitute for raw attributes. These metrics act as a "warm start" for the model. For instance, in the TUDataset benchmarks, Christopher Morris and Nils M. Kriege highlight that simple local statistics like node degree are often sufficient for GNNs to achieve competitive results on unattributed graphs.

In more complex scenarios, such as combinatorial optimization, Frederik Wenkel suggests using a "geometric foundation" of eccentricity and triangle counts to enable the network to learn topological patterns relevant to the task at hand.

Learning the "Rules" of Connectivity

Beyond simple statistics, recent work focuses on learning the underlying generative rules of a graph's evolution. Hisashi Kashima and Naoki Abe pioneered this by modeling network growth as a copy-and-paste evolutionary process, allowing for link prediction without any node metadata.

This "rule-based" approach has evolved into inductive structural learning. For example, Jaejun Lee and Joyce Jiyoung Whang developed InGram, which predicts missing facts in knowledge graphs by building a "relation graph." This allows the model to generalize to entirely new entities and relations based solely on how they fit into the existing topological web.

Capturing Higher-Order Invariants

A major hurdle in unattributed learning is that local connectivity (like degree) can be "noisy." To solve this, researchers are looking at higher-order patterns:

Why It Works: The Topology is the Signal

The success of these methods suggests a fundamental shift in how we view graph data. In many systems, the topology is the primary signal, and attributes are merely secondary annotations.

As Santiago Segarra and Antonio G. Marques demonstrate, we can even recover missing node features by mapping nodes into a structural embedding space and "borrowing" features from topologically similar nodes. This implies that in a well-connected graph, a node's identity is defined not by what it is, but by where it sits in the network. By prioritizing coordinate-free structural learning and topological invariants, we can build models that are robust to missing data and capable of generalizing across the vast, unattributed landscapes of real-world relational data.

Papers in this version

Showing 60 of 82 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_learning_from_pure_topology_in_unattribu_6a1316a3,
  title = {Learning from Pure Topology in Unattributed Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-from-pure-topology-in-unattributed-graphs/rdr_6a1316a344a84c40aa095730af9f79fe},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_from_pure_topology_in_unattribu_6a1316a3_v1,
  title = {Learning from Pure Topology in Unattributed Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-from-pure-topology-in-unattributed-graphs/rdr_6a1316a344a84c40aa095730af9f79fe/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}