In the standard narrative of Graph Neural Networks (GNNs), nodes are defined by two distinct types of information: their attributes (like a user's profile data) and their structure (who that user is connected to). Most models treat structure as a routing mechanism—a way to aggregate and smooth these attributes across the network.
However, a growing body of research suggests that this hierarchy is often inverted. In many real-world scenarios, node attributes are missing, noisy, or entirely redundant. In these cases, the "shape" of the network—its topology—is not just a conduit for information, but the primary signal itself. Researchers are finding that by treating structural properties as surrogates for missing features, we can build models that are more robust, efficient, and sometimes even more accurate than those relying on rich metadata.
The Geometry of Missing Information
The most immediate motivation for this direction is the "attribute-missing" problem. In practical applications, data is rarely complete; users hide profiles, or sensors fail. Wenxuan Tu and Sihang Zhou demonstrate that when features are missing, the graph's topology serves as a superior "anchor" for imputation compared to standard statistical methods like mean-filling.
This works because the connectivity of a node is rarely random. Chi Thang Duong and Karl Aberer show that structural information derived from topology (such as DeepWalk embeddings) can serve as a high-quality substitute for raw attributes. In some datasets, these "artificial" features actually outperform the original "real" features, suggesting that the inherent geometry of the network often contains a cleaner signal for classification than the noisy metadata we typically collect.
Synthesizing Features from Structure
If the structure is the signal, how do we represent it for a neural network? The simplest approach is to engineer "synthetic" features from the graph itself.
- Local Statistics: Yanqiao Zhu and Lifang He use local topological statistics—like node degrees or clustering coefficients—as surrogates for biological features in brain networks.
- Path Distances: Qingju Jiao and Peige Zhao propose $t$-hopGCN, which uses shortest-path distances between nodes as features. This "plug-and-play" lift allows standard GNNs to operate on featureless graphs by treating the relative position of nodes as their primary identity.
- Global Distributions: Raul Barbosa and colleagues take this further with "weightless" architectures that classify graphs based solely on the statistical distributions of node degrees and centrality.
Topological Data Analysis and Persistence
Beyond simple statistics, researchers are looking at the "global shape" of graphs through the lens of Topological Data Analysis (TDA). The intuition here is that graphs have multi-scale structures—loops, holes, and clusters—that persist even when local details change.
Bastian Rieck and Karsten Borgwardt introduced a persistent Weisfeiler-Lehman procedure that infuses traditional graph kernels with persistent homology. By tracking how topological features appear and disappear as we "filter" the graph (e.g., by node degree), they create a signature of the graph's structure that is far more descriptive than simple neighbor aggregation. Similarly, Qi Zhao and Yusu Wang developed frameworks to "learn" which of these topological summaries matter most, often outperforming complex GNNs using only these structural "skeletons."
Why Structure is Often Enough
A central question in this research is why models can succeed even when node features are replaced with random noise. Yaming Yang and Ziyu Guan demonstrate that in many tasks, the arrangement of connections encodes more critical information than the attributes themselves. This is particularly true in cases of automorphic equivalence, where nodes play identical structural roles. Fengli Xu and Yong Li argue that capturing these structural symmetries provides a robust learning signal that remains effective even in the complete absence of external data.
This shift in perspective is leading to a new taxonomy of datasets. Renming Liu and Anna Little have shown that many benchmark tasks are "purely structural." In these cases, the original node features are essentially a distraction; replacing them with basic topological metadata like node degrees can actually boost performance.
Efficiency and Interpretability
Finally, prioritizing structure over high-dimensional features leads to massive gains in efficiency. Astrit Tola and Baris Coskunuzer developed TopER, which reduces complex topological evolution into a simple 2D representation. By focusing on how the graph "grows" rather than the minutiae of node attributes, they achieve state-of-the-art accuracy while being orders of magnitude faster than standard deep learning approaches.
The intuition across all these works is clear: the network is not just a container for data; the network is the data. By treating topology as a first-class citizen, we can build models that are not only more resilient to missing information but also more deeply attuned to the fundamental patterns of connectivity that define complex systems.