Correlated Binary Data for Machine Learning

Data sets containing instances that are assigned values by an ensemble of annotators of unknown accuracy are becoming increasingly common. Binary, potentially correlated data are frequent in a number of disciplines, and thus eligible to be exploited by ensemble meta-learners. A prior key step is testing the meta-learners with synthetic data sets featuring realistic correlation patterns, which is the main scope of this work. To achieve this goal, two challenges are faced: (i) finding out a new correlated pattern to model Bernoulli random variables, and (ii) obtaining a process to generate realistic synthetic data sets. A comparative analysis and performance results are provided for two methods of artificial data generation. The methods are also tested using two state-of-the-art binary ensemble meta-learners that consider inter-classifier dependencies.