bayeclass: an R package for learning Bayesian network classifiers. Applications to neuroscience

Machine learning provides tools for automatized analysis of data. The most commonly used form of machine learning is supervised classification. A supervised classifier learns a mapping from the descriptive features of an object to the set of possible classes, from a set of features-class pairs. Once learned, it is used to predict the class for novel data instances. The Bayesian network-based supervised classifiers are particularly useful. They have a solid theoretical basis in probability theory and provide competitive predictive performance. Many algorithms for learning Bayesian network classifiers exist. However, only two are provided in freely available software. The R software environment is the leading open-source system for statistical computing. We provide an implementation of state-of-the-art Bayesian network classifiers for the R environment, in the form of an add-on package called bayesclass. The best-known Bayesian network classifier is the naive Bayes. It assumes that the features are independent given the class. In many domains, more accurate classification is obtained by relaxing these assumptions. The semi-naive Bayes model removes all assumptions of conditional independence within disjoint subsets of features. Its state-of-the-art learning algorithm, the backward sequential elimination and joining (BSEJ) algorithm, tends to produce semi-naive Bayes models with small subsets of related features. Such a model removes only a few of naive Bayes’ independence assumptions. We extend the BSEJ algorithm with a second step which removes some of its unwarranted independence assumptions. Our classifier outperforms the BSEJ and five other Bayesian network classifiers on a set of benchmark databases, although the difference in performance is not statistically significant. Classification of neurons is an important problem in neuroscience. Previously, 42 experts classified a set of interneurons according to a proposed taxonomy. They disagreed on many of the terms of the taxonomy. The gathered data allows for constructing an objective computational model that could resolve the conflicts by assigning the definitive class labels. A supervised classifier can learn the required mapping from quantitative neuronal descriptors to the experts’ taxonomical choices. A challenge is that there are up to 42 class labels per neuron and the most voted label is not always reliable (i.e. has many votes). We use only the cells with reliable class labels to build the classifiers and obtain high predictive accuracy.