50. Using Information Gain to Compare the Effeciency of Machine Learning Techniques When Classifying Influenza Based on Viral Hosts

The paper compares the performance of two classical machine learning techniques when features selection is used to improve Influenza-A host classification. The impact of using the most informative positions on both the classifier efficiency and performance of decision trees (DTs) and neural networks (NNs) was measured. The experiments were conducted on cDNA sequences belonging to all the viral segments of the subtype H1 to ensure authenticity of results. Sequences belonging to each viral segment were further divided into viruses infecting human and non-human hosts prior to classification analysis. The performance measures, accuracy, sensitivity, specificity, precision, and time, were used. Extracting the best hundred informative positions with the information gain (IG) algorithm increased classification efficiency for both classifiers by more than 80 % for all viral segments. The change in performance was insignificant. The overall results of statistical significant tests showed that NNs classified viral hosts more accurately than DTs for subtype H1. The tests also showed that the DTs are significantly faster than NNs in classifying Influenza hosts despite the slight decrease in performance.