Zhu and Wang's NINL initializes each node with the sum of original-graph degrees in its closed radius-r neighborhood. The paper sets r to the ceiling of the graph's average shortest-path length. Each iteration then replaces every node's score by the sum of its neighbors' previous scores: NINL-p = A^p NINL-0. The paper uses p = 3; zero iterations returns the initial degree volume. Repeated vertices and edges in these walks count.
centrality_ninl(x, ninl_order = 3, ninl_radius = NULL, ...)Named numeric vector in input node order.
Network input accepted by centrality.
Nonnegative integer iteration count, default 3.
At most 2^53 - 1, the consecutive-integer precision of doubles.
NULL for the source-defined automatic radius,
or a nonnegative integer hop radius, or Inf for all reachable
nodes. Radius zero uses the focal node's degree alone.
Additional arguments to centrality.
normalized = TRUE divides by the maximum score; all-zero scores
remain zero. Normalization is optional and is not part of the raw
definition in the original paper.
Uses simple undirected unweighted topology: either arc creates an edge; loops and parallel edges are removed. Weights, mode, inversion and cutoff are ignored. This does not claim a directed or weighted NINL definition.
The mean path length includes all distinct vertex pairs. For disconnected graphs it is infinite, so the automatic radius includes every reachable node in each component. This is an explicit cograph extension of the paper's connected example; unreachable nodes never enter the degree sum. Isolates score zero and empty graphs return no scores. A supplied radius is an explicit generalization of the paper's automatic-radius rule.
Stepwise propagation evaluates the requested finite iteration count, without assuming convergence to eigenvector centrality. Exact repeated floating-point states of period one or two allow the remaining iterations to be skipped while preserving parity. No tolerance-based convergence cutoff is used. Normalized scores can alternate on bipartite graphs. Dense distance calculation and propagation take O(n cubed + p n squared) time and O(n squared) memory; very large orders can be slow if no exact repeated state occurs. Raw overflow raises an error. With maximum normalization, global rescaling after every step avoids overflow; extremely small relative scores can still underflow in double precision.
Zhu, J. and Wang, L. (2021). Identifying Influential Nodes in Complex Networks Based on Node Itself and Neighbor Layer Information. Symmetry, 13, 1570. tools:::Rd_expr_doi("10.3390/sym13091570").
centrality_ninl(igraph::make_graph("Zachary"))
centrality_ninl(igraph::make_star(5, mode = "undirected"), ninl_order = 2)
Run the code above in your browser using DataLab