Air Transport Network: A Comparison of Statistical Backbone Filtering Techniques
Résumé
The big break in data collection tools of large-scale networks from biological, social, and technological domains expands the challenge of their visualization and processing. Numerous structural and statistical backbone extraction techniques aim to reduce the network’s size while preserving its gist. Here, we perform an experimental comparison of seven main statistical methods in an air transportation case study. Correlations analysis shows that Marginal Likelihood Filter (MLF), Locally Adaptive Network Sparsification Filter (LANS), and Disparity Filter are biased toward high weighted edges. We compare the extracted backbones using four indicators: the size of the largest component, the number of nodes, edges, and the total weight. Results show that techniques based on a binomial distribution null model (MLF and Noise Corrected Filter) tend to retain many edges. Conversely, Disparity Filter, Polya Urn Filter, LANS Filter, and Global Statistical Significance Filter (GLOSS) are pretty aggressive in filtering edges. The ECM Filter lies between these two behaviors. These results may guide users in selecting appropriate techniques for their applications.