The effect of the hyperparameters of an RBF SVM on classifying pneumonia

Sammanfattning: Pneumonia is a deadly disease if gone untreated. In order to treat it, first it has to be discovered, which can be done by examining patterns of chest X-ray images. This process can be done by human experts but there is also a desire to make it more efficient by utilizing machine learning. One machine learning model well equipped for handling classification tasks is the Support Vector Machine (SVM) with a Radial Basis Function (RBF) Kernel. However, there exists a vast amount of configurations which greatly affects the achieved accuracy and therefore there is a need of understanding how the different configurations work and how they affect the behavior of the SVM. Therefore we aimed to answer specifically how the two hyperparameters cost and gamma affect the performance of an RBF kerneled SVM when using Histogram of Oriented Gradients (HOG) of chest X-ray images as features. This investigation was conducted by performing a coarse grid search combined with cross-validation with SVM as estimator. The accuracies obtained were then displayed on a heatmap and afterwards compared. In our results, it was found that there does exist a clear correlation between cost, gamma and the achieved accuracy. In some areas of the hyperparameter space, the parameters were able to compensate for each other, i.e., if one of them increased, then the other one needed to decrease in order to retain its accuracy. There also existed anomalies to the patterns discerned, when, in some extreme values for either hyperparameter, the effect of the other disappeared, implying an independency. Lastly, the best accuracies could be found in the middle region of the hyperparameter space.