Abstract: Most measures of the structural and topological properties of a complex network are of a combinatorial nature, basically counting certain nodes, edges, triangles, paths, etc. But the typically finite set of nodes of the studied network is often either explicitly or implicitly considered representative of a much larger finite or infinite set of objects of interest, either by being a subset of this larger set or by being some kind of discretisation or aggregation of it. Examples are networks consisting of a statistically sampled number of members of a large social or scientific community, networks of discrete regular grid points on a surface or of irregular mesh cells in some manifold, or networks constructed from recurrence plots of time series with regular or irregular time intervals. This selection procedure typically results in individual nodes of the studied network representing quite differently sized regions of the domain of interest, inducing substantial biases in derived network statistics.
Introduction: The Problem of Representation Bias
In classical graph theory, every node is treated as a discrete, equal unit. However, in many real-world applications, nodes serve as proxies for parts of a larger continuous domain. When these nodes represent regions of varying size—such as grid cells on a globe or brain regions of different volumes—standard network metrics can produce misleading results.
Consider a climate network mapped onto a latitude-longitude grid. Because grid lines converge at the poles, a single degree of longitude represents a much smaller physical area at the North Pole than at the Equator. Standard calculations of "node degree" (the number of connections) will erroneously identify the poles as massive hubs simply because more nodes are packed into a smaller physical space. This is known as representation bias. Similar issues affect neuroscience, where brain regions (Regions of Interest, or ROIs) vary significantly in volume, and economics, where countries have vastly different populations.
The n.s.i. Framework: Redefining Network Metrics
To address these biases, this research introduces a robust axiomatic framework based on the concept of Node Splitting Invariance (n.s.i.). The core principle is that if a node is split into two "twins" that share the same neighbors and divide the original node's "weight" (e.g., area or volume), the resulting network measure should remain unchanged.
By applying this principle, common metrics are redefined:
- Node-Weighted Degree: Instead of counting neighbors, the weights of those neighbors are summed.
- Clustering Coefficient: Redefined as the probability that two "weight units" linked to a node are also linked to each other.
- Betweenness Centrality: Modified to account for the weights of nodes along a path, ensuring that paths through many small nodes are not prioritized over paths through a single large node.
Empirical Results and Applications
Climate Networks
The impact of the n.s.i. framework is most visible in global climate networks. In standard unweighted networks, the poles appear to have artificially high degree and clustering. When n.s.i. measures are applied—weighting nodes by the cosine of their latitude to account for the Earth's spherical geometry—these artifacts vanish. This reveals that the perceived "hubs" at the poles were merely a result of grid geometry rather than underlying climate physics.
Neuroscience and Social Networks
In the human brain network, accounting for ROI volume significantly shifts the rankings of regional importance. For instance, the right thalamus drops significantly in clustering rank when its relatively small volume is accounted for.
Similarly, in the internet's Autonomous Systems (AS) network, standard clustering coefficients are low (~0.21), but the node-weighted version (weighting by controlled IP address space) is much higher (~0.8). This suggests that while the nodes appear sparsely connected, the actual IP address space is highly integrated, correcting a common "small-world" illusion.
Conclusion
The n.s.i. property serves as a vital "sanity check" for network science. If a metric changes simply because a node is subdivided without changing the underlying connectivity, the metric is measuring the sampling method rather than the system itself. While choosing the appropriate weight (GDP, population, area, etc.) remains a research-specific decision, this framework provides the mathematical consistency necessary to analyze networks that represent continuous or aggregated domains accurately.
Summary updated