Using genetic algorithms versus line search optimization for HIV predictions

Neural Networks are used as pattern recognition tools in data mining to classify HIV status of individuals based on demographic and socio-economic characteristics. The data consists of seroprevalence survey information and contains variables such as age, education, location, race, parity and gravidity. The multilayer perceptron (MLP) neural network architecture was used for this study since as preliminary design showed this architecture to be the most optimal. The design of classifiers involves the assessment of classification performance, and this is based on the accuracy of the prediction using the confusion matrix. Three design approaches were implemented and a comparative analysis done. An accuracy of 84.24% was obtained for the genetic algorithms meanwhile an accuracy of 74% is obtained for the standard optimized network. The network structures for the different methodologies as well as the training and optimization times are also different. The gradient method proved to be the less computationally expensive but the most erroneous. A committee of neural networks was also analyzed and this obtained the same accuracy as the genetic algorithm network thus emphasizing the genetic algorithm as the better method.