Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Layer 1 Layer 2 Layer 3 Layer 4
Interconnected EconomiesEugene Neduv, Xiao Xu, Kehan Lu IEOR, School of Engineering and Applied Science
US Equity Market: InterconnecIons between largest market cap US companies
References1. “The mul+plex dependency structure of financial markets”, Nicol o Musmeci, Vincenzo Nicosia, Tomaso Aste, Tiziana Di MaDeo,and Vito Latora2. “Centrality in Interconnected Mul+layer Networks” Manlio De Domenico, Albert Sol e-Ribalta, Elisa Omodei, Sergio G omez,and Alex Arenas3. “A Graph-theore+c perspec+ve on centrality” Stephen P. BorgaV , Mar+n G. EvereD
Multilayered Networks: Layers represent various types of relationshipsInformation carried by each layer is sufficiently different: correlation between layers and unique edgeweight quantify the differences . Interlayer links weights define relative importance of information bytype. Tensor algebra works for multiplex networks measures. Node degree – number of neighbors,weighted degree sum of neighboring edge weights are basic node measures.
Further ResearchClustering of companies based on mul+ple layers of connec+ons. Iden+fy groups with similarcharacteris+cs. Can be used as a recommenda+on engine for porZolio selec+on
Centrality: Represents node influence in the systemCentrality types: Weighted Degree, PageRank, Closeness, Betweenness.Chose appropriate centrality for the process being modeled.
• Strength Tensor • Eigenvector Centrality
• DiffusionEquaIon
• Laplacian
Trends in business dominance over the past 10 years
Correlation Centrality by Sector
MulIplex Centrality by Sector CompeIIon Centrality by Sector
Supply-Customer Centrality by Sector
Information diffusion can be generalized to multilayered network to account for interlayer diffusion.
Interlayer Degree Correlation!" = $"
%&%
'[%,"] =∑,(!,
% − ! % )(!," − ! " )
∑,(!,% − ! % )0(!,
" − ! " )0
Fraction of Edges unique to Layers
1[%] =2
03 %4,,5
6,5% 7
"8%
(2 − 6,5["])
Average Edge Overlap
9 =203
4,,5
46,5[%]
Degree Evolution by Layer
:;"<=:>
= ?"<=%<@;%<@ > −?"<=
%<@ 1%<@A'<B;"<= >:;"<=:>
= −C"<=%<@;%<@(D) ;"<=(>) = ;%<@(E)F
GC"<=%<@>
Random walks on mul+layered network include interlayer jumps.H"<= > + 2 = −C"<=
%<@H%<@ D , J"<=%<@ = ="<=
%<@ − C"<=%<@
Eigenvectors of Laplacian tensor can be viewed as a centrality measure. Unfolding of a tensor into asupra matrix generalizes eigenvector problem.
C"<=%<@K%<@ = LK"<=
K"<= = M2G2?"<=
%<@K%<@• Bonacich Eigenvector Centrality
K%<@• Unfolded rank-1 supra eigenvector
Using tensor algebra all monoplex measures can be generalized to mul+layered networks.Studying spectral proper+es of graph Laplacian can be used to understand the dynamics ofinforma+on diffusion.
Acknowledgments DSI Summer Research Funding, Lab for Intelligent Asset Allocation