Skip to content
Code accompanying the paper "Three hypergraph eigenvector centralities."
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Hypergraph eigenvector centralities

This code and data repository accompanies the paper

  • Three hypergraph eigenvector centralities. Austin R. Benson. arXiv:1807.09644, 2018.

All of the code is written in Julia 1.0.

For questions, please email Austin at


The tags and DAWN datasets are in the data/ directory. These files were downloaded directly from here and here. The ngrams dataset is available from, but I cannot include the data directly due to the usage terms. The script data/convert_ngrams.jl can be used to convert the raw ngrams data into a format that can be automatically read by the existing code.


First, download the code:

git clone

The code is written in Julia 1.0. To reproduce everything, you will need the following packages:

using Pkg

Then you can test the code


Examples for computing eigenvectors

# Create the hypergraph adjacency tensor for the hypergraph in Proposition 2.8.
# Hyperedges: {{1, 2, 3}, {1, 2, 4}, {3, 5, 6}, {5, 6, 7}}
T = SymTensor3([1, 1, 3, 5], [2, 2, 5, 6], [3, 4, 6, 7], ones(Float64, 4))

# We might first want to take the largest component.
Tlcc, lcc_inds = largest_component(T)
# However, the hypergraph is connected here.

# Clique motif Eigenvector Centrality (CEC)
(cec_c, cec_converged) = CEC(T)
# Z-eigenvector Centrality (ZEC)
(zec_c, zec_converged) = ZEC(T)
# H-eigenvector Centrality (HEC)
(hec_c, hec_converged) = HEC(T)

Computational experiments for Proposition 2.8

Proposition 2.8 says that Z-eigenvector centrality (ZEC) vectors can be unstable in the sense of the definitions the paper from Kolda and Mayo analyzing their symmetric shifted higher-order power method for computing Z-eigenvectors of symmetric tensors. The consequence is that we cannot rely on power-method like algorithms for computing ZEC vectors. This script contains the computational experiments demonstrating the proof of the proposition:


The output should look something like the following

dynamical systems algorithm converged: true
tensor z-eigenvalue: 1.4142111552433783
tensor z-eigenvector (2-norm normalized): [0.408248, 0.408248, 0.471403, 0.235704, 0.408248, 0.408248, 0.235704]
orthogonality test (should be close to 0): 9.71445146547012e-16
orthonormality test (should be close to 0): [3.70907e-16, 2.66455e-16, 2.35813e-16, 2.87556e-16, 1.0074e-15, 9.14796e-16]
Spectrum of projected Hessian of the Lagrangian: [-2.82842, -2.82842, -2.82842, -2.06111, -1.41421, 0.646902]

Notice that the projected Hessian is indefinite since it has both a negative and positive eigenvalue.

Computational experiments for sunflower hypergraphs.

The paper provides analytic results for sunflower hypergraphs where the core is a singleton. The script sunflower.jl provides numerical evidence for both the results listed in the right of Figure 1 and Proposition 2.12.

sf_test1()  # computational experiments related to the table within Figure 1.
sf_test2()  # computational experiments related to Proposition 2.12.

Reproduce the figures and tables in the paper

First, we need to compute all of the centralities. Here is an exmaple for the 4-uniform tags dataset. This code is designed to use the data format of the temporal higher-order networks available here.

collect_data("tags-ask-ubuntu", 4, false)  # produces tags-ask-ubuntu-4.jld2
collect_data("DAWN", 5, true)  # produces DAWN-5.jld2

In the second command, the parameter "4" means 4-uniform hypergraph, and the parameter "false" means to not look for exact size matching in the dataset. In this case, A size-5 hyperedge becomes five different size-4 hyperedges. In the third command, we look at a 5-uniform hypergraph, and we only take hyperedges in the original data since the third parameter is set to "true".

We have stored all of the computational results in the results/ directory:

  • results/ngrams-{3,4,5}.jld2
  • results/tags-ask-ubuntu-{3,4,5}.jld2
  • results/DAWN-{3,4,5}.jld2
Figures 2, 3, and 4.

These are the rank correlation plots.


# Create ngrams-{3,4,5}-rankcorr.eps (x axis starts from 20)
for k in [3, 4, 5]; rank_corr("ngrams", k, 20); end 

# Create tags-ask-ubuntu-{3,4,5}-rankcorr.eps (x axis starts from 10)
for k in [3, 4, 5]; rank_corr("tags-ask-ubuntu", k, 10); end

# Create DAWN-{3,4,5}-rankcorr.eps (x axis starts from 10)
for k in [3, 4, 5]; rank_corr("DAWN", k, 10); end
Table 1

This table has the summary statistics of the datasets.

for k in [3, 4, 5]; summary_statistics("ngrams", k); end
for k in [3, 4, 5]; summary_statistics("DAWN", k); end
for k in [3, 4, 5]; summary_statistics("tags-ask-ubuntu", k); end
Tables 2, 3, and 4

These are the top-k ranked elements according to the centrality vectors.


# Table 2 
# Print top-20 words in ngrams datasets in a single row
top_ranked("ngrams", 20, true)

# Table 3
# Print top-10 tags in ask ubuntu dataset; one block for each hypergraph uniformity.
top_ranked("tags-ask-ubuntu", 10, false)

# Table 4
# Print top-10 drugs in DAWN dataset; one block for each hypergraph uniformity.
top_ranked("DAWN", 10, false)
You can’t perform that action at this time.