Greedy Search among Decomposable Models

Vladislav Bína


Learning Algorithms and Training, Machine Learning, Greedy Search, Compositional Models, Decomposable Models, Decomposable Likelihood-ratio


The contribution presents a proposal of a greedy method for learning in context of compositional models, i.e. in the class of probabilistic models used for representation of multidimensional probabilistic distributions. For this task a sequence of low--dimensional distribution is used which is joint together with the operator of composition. The greedy method takes advantage of properties of compositional models and performs a suboptimal search among the subclass of decomposable (compositional) models. The search algorithm is based on likelihood--ratio test which controls the exploration of structure of the conditional independence relations among model variables. The main sphere of application stems from the methodology; the method is advantageous for efficient representation and handling of multidimensional distribution with categorial variables (variable counts in order of tens).

Important Links:

Go Back