After creating a graph, conditional probability distribution parameters can be calibrated using historical data or reflecting expert judgement.


  • Several optimization techniques available.

  • Start simple with regression under linear assumptions on node dependencies.

  • Incorporate expert judgment in parameterizing nodes distributions.

  • Data imputation techniques available for incomplete datasets.

  • Frequency matching.

  • Preprocessing methods available prior to calibration.

  • Data transformation routines. Node types: Categorical, Discrete, Continuous, Mixture, Deterministic.


[1]  Koller D, Friedman N. Probabilistic graphical models: principles and techniques. MIT press; 2009.  - III Learning  - 17. Parameter Estimation p717

[2]  Sucar LE. Probabilistic graphical models. Advances in Computer Vision and Pattern Recognition. London: Springer London. doi. 2015;10(978):1.

[3]  Darwiche A. Modeling and reasoning with Bayesian networks. Cambridge university press; 2009 Apr 6. - 17 Learning: The Maximum Likelihood Approach p439, 18 Learning: The Bayesian Approach p477

© 2020 by Tenokonda Ltd. All rights reserved.