A Particle Swarm Optimization approach to mixed attribute data-set classification

We describe a Particle Swarm Optimization (PSO) approach to the problem of classifying mixed-attribute data sets. It relies on retrieving optimal particle positions in the search space that correspond to the centroids of classes. When evaluating the fitness function, we use different mechanisms to interpret the particle positions in the description space, based on data type; as will be described, rounding is used for integer attributes while a frequency measure is used for categorical descriptors. An experimental set up was realized and tested on the Adult database, leading to recognition accuracies that were better than those obtained with well known classifiers.