TY - GEN

T1 - Approximate conditional independence test using residuals

AU - Uda, Shinsuke

N1 - Funding Information:
This work was supported by the Creating information utilization platform by integrating mathematical and information sciences, and development to society, CREST (JPMJCR1912) from the Japan Science and Technology Agency (JST) and by the Japan Society for the Promotion of Science (JSPS) KAK-ENHI Grant Number (JP18H04801, JP18H02431) and Kayamori Foundation of Informational Science Advancement.

PY - 2020

Y1 - 2020

N2 - Conditional mutual information is a useful measure for detecting the association between variables that are also affected by other variables. Though permutation tests are used to check whether the conditional mutual information is zero to indicate mutual independence, permutations are difficult to perform because the other variables in a dataset may be associated with the variables in question. This problem is particularly acute when working with datasets of small sample size. This study aims to propose a computational method for approximating conditional mutual information based on the distribution of residuals in regression models. The proposed method can implement the permutation tests for statistical significance by translating the problem of measuring conditional independence into the problem of estimating simple independence. Additionally, a reliability of p-value in permutation test is defined to omit unreliably detected associations. We tested our proposed method's performance in inferring the network structure of an artificial gene network against comparable methods submitted to the Dream4 challenge.

AB - Conditional mutual information is a useful measure for detecting the association between variables that are also affected by other variables. Though permutation tests are used to check whether the conditional mutual information is zero to indicate mutual independence, permutations are difficult to perform because the other variables in a dataset may be associated with the variables in question. This problem is particularly acute when working with datasets of small sample size. This study aims to propose a computational method for approximating conditional mutual information based on the distribution of residuals in regression models. The proposed method can implement the permutation tests for statistical significance by translating the problem of measuring conditional independence into the problem of estimating simple independence. Additionally, a reliability of p-value in permutation test is defined to omit unreliably detected associations. We tested our proposed method's performance in inferring the network structure of an artificial gene network against comparable methods submitted to the Dream4 challenge.

UR - http://www.scopus.com/inward/record.url?scp=85083109212&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85083109212&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:85083109212

T3 - ICAART 2020 - Proceedings of the 12th International Conference on Agents and Artificial Intelligence

SP - 297

EP - 304

BT - ICAART 2020 - Proceedings of the 12th International Conference on Agents and Artificial Intelligence

A2 - Rocha, Ana

A2 - Steels, Luc

A2 - van den Herik, Jaap

PB - SciTePress

T2 - 12th International Conference on Agents and Artificial Intelligence, ICAART 2020

Y2 - 22 February 2020 through 24 February 2020

ER -