TY - GEN
T1 - Directed network topology inference via sparse joint diagonalization
AU - Shen, Yanning
AU - Fu, Xiao
AU - Giannakis, Georgios B.
AU - Sidiropoulos, Nicholas D.
PY - 2018/4/10
Y1 - 2018/4/10
N2 - Discovering the connectivity patterns of directed networks is a crucial step towards understanding complex systems such as human brains and financial markets. Network inference approaches aim at estimating the hidden topology given nodal observations. Existing approaches relying on structural equation models (SEMs) require full knowledge of exogenous inputs, which may be unrealistic in certain applications. Recent tensor-based alternatives advocate reformulation of SEMs as a three-way tensor decomposition task that only requires second-order statistics of exogenous inputs for identifying the hidden topology. However, the tensor-based methods are computationally expensive, and is hard to incorporate prior information of the network structure (e.g., sparsity and local smoothness), but prior information is often important for enhancing performance. The present work puts forth a joint diagonalizaition (JD)-based formulation for directed network inference. JD can be viewed as a variant of tensor decomposition, but features more efficient algorithms and can readily incorporate prior information of network topology. New topology identification guarantees that do not rely on knowledge of exogenous inputs are established. Judiciously designed simulations are presented to showcase the effectiveness of the proposed approach.
AB - Discovering the connectivity patterns of directed networks is a crucial step towards understanding complex systems such as human brains and financial markets. Network inference approaches aim at estimating the hidden topology given nodal observations. Existing approaches relying on structural equation models (SEMs) require full knowledge of exogenous inputs, which may be unrealistic in certain applications. Recent tensor-based alternatives advocate reformulation of SEMs as a three-way tensor decomposition task that only requires second-order statistics of exogenous inputs for identifying the hidden topology. However, the tensor-based methods are computationally expensive, and is hard to incorporate prior information of the network structure (e.g., sparsity and local smoothness), but prior information is often important for enhancing performance. The present work puts forth a joint diagonalizaition (JD)-based formulation for directed network inference. JD can be viewed as a variant of tensor decomposition, but features more efficient algorithms and can readily incorporate prior information of network topology. New topology identification guarantees that do not rely on knowledge of exogenous inputs are established. Judiciously designed simulations are presented to showcase the effectiveness of the proposed approach.
KW - CANDE-COMP/PARAFAC (CP) decomposition
KW - Structural equation models
KW - joint diagonalization
KW - network topology inference
UR - http://www.scopus.com/inward/record.url?scp=85050962392&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85050962392&partnerID=8YFLogxK
U2 - 10.1109/ACSSC.2017.8335433
DO - 10.1109/ACSSC.2017.8335433
M3 - Conference contribution
AN - SCOPUS:85050962392
T3 - Conference Record of 51st Asilomar Conference on Signals, Systems and Computers, ACSSC 2017
SP - 698
EP - 702
BT - Conference Record of 51st Asilomar Conference on Signals, Systems and Computers, ACSSC 2017
A2 - Matthews, Michael B.
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 51st Asilomar Conference on Signals, Systems and Computers, ACSSC 2017
Y2 - 29 October 2017 through 1 November 2017
ER -