Characterizing Bayesian network structure sensitivity in classifiers derived from data.

摘要

Bayesian networks are used to model causal relationships in which the network is composed of a directed acyclic graph (DAG) with corresponding probability distributions for each node. I investigated the use of DAG enumeration as a method of evaluating differences in network structures inferred for individuals when the number of variables is small. For computational efficiency, we used a DAG list that relates each DAG to a specific equivalence class via a variable permutation, which reduces classification essentially to a table look-up. Additionally, we mapped the enumeration results to a circle to allow visualization of the entire hypothesis space in one graph. To evaluate our approach, we applied our methodology to data from the field of sleep and circadian rhythms. Our results suggest that this approach is an effective method of data analysis for inferring relationships among variables when the number of variables is small.