Knowledge discovery in data and causal models in analytical informatics

• Дата: 2018
• Автор: Balabanov, O.S.
• Формат: Стаття
• Мова: Ukrainian
• Опубліковано: Інститут програмних систем НАН України 2018
• Теми: causal network; model inference from data; Markov properties; conditional independence; structure of dependencies; causal effect; edge orientation; d-separation; UDC 004.855:519.216
• Онлайн доступ: https://pp.isofts.kiev.ua/index.php/ojs1/article/view/299
• Назва журналу: Problems in programming

The methodology of inductive inference of causal models is briefly overviewed. We argue that causal networks, being recovered from data, are able to describe adequately a structure of influences in environment (object) at hand. It’s a causal model that is required when predicting the effect of intervention in object. We outlined the preconditions and requirements on data collection process in aiming to reach an adequate causal network. A multivariate statistical data sample (measured under unified scheme) is needed in the input of inference method. We consider an independence-based approach to causal inference. Methods of this approach are correct, and can perform well in presence of hidden variables. The method’s output usually contains some edges not exactly oriented. Uncertainty of such kind is predetermined by problem setting and allows retaining model adequacy. We suggest a way to enforce an inference algorithm due to set of resolutions which reduce a space for searching separating sets (so focusing a process of edge verification). The modification proposed is based on systematic utilization of concept of locally–minimal separating set and Markov properties. An efficiency of developed algorithms (‘Razor’ series) is demonstrated by control experiments and case study. A distinction between a prediction of causal effect (i.e. effect of active experiment) and traditional prediction in data analysis is illuminated. Some problems of parameter estimation are presented. Some opportunities to predict causal effect when model is incompletely identified are illustrated. We point out a few ideas and new research trends which can enrich analyst’s ability to verify or identify a model. Problems in programming 2017; 3: 96-112