Efficient implementation of compositional models for data mining

0497540 - ÚTIA 2019 RIV JP eng C - Conference Paper (international conference)
Kratochvíl, Václav - Jiroušek, Radim - Lee, T. R.
Efficient implementation of compositional models for data mining.
Proceedings of the 21st Czech-Japan Seminar od Data Analysis and Decision Making. Japan: Aoyama Gakuin University, Japan, 2018 - (Sung, S.; Vlach, M.), s. 80-87. ISBN 978-80-7464-932-5.
[The 21st Czech-Japan Seminar on Data Analysis and Decision Making. Kamakura (JP), 23.11.2018-26.11.2018]
R&D Projects: GA ČR(CZ) GA16-12010S
Grant - others:AV ČR(CZ) MOST-18-04
Program: Bilaterální spolupráce
Institutional support: RVO:67985556
Keywords : data mining * mutual information * compositional models * conditional independence * probability theory
OECD category: Automation and control systems
http://library.utia.cas.cz/separaty/2018/MTR/kratochvil-0497540.pdf

A compositional model encodes probabilistic relationships among variables of interest. In connection with various statistical techniques, it represents a practical tool for data modeling and data mining. Structure of the model represents (un)conditional independencies among all variables. Relationships of dependent variables are described by low-dimensional probability distributions. Having a compositional model, a data miner can easily apply an intervention on variables of interest, fix values of other variables (conditioning), or to narrow the context of a problem (marginalization). The model learning process can be controlled to avoid overfitting of data.

In this paper, we present a new semi-supervised web application that will enable researchers to design probabilistic (compositional) models (both causal and stochastic). Thanks to the web architecture of the system, the researchers will always have a possibility to influence the data-based model construction process from any place of the world. It is also expected that the application of this methodology to practical problems will open new problems that will be an inspiration for further theoretical research.
Permanent Link: http://hdl.handle.net/11104/0291220