Machine learning approaches for biomaterial modeling

In this thesis, we explore the use of different machine learning approaches for predicting the amount of protein adsorption on biomaterial surfaces. The models built using machine learning approaches would be a valuable tool for biomaterial modeling. Feature selection is very important for this problem domain. Several filter based and wrapper based feature selection methods were tried. The wrapper based approach with a genetic algorithm gives the best results. Machine learning schemes like model trees, regression trees, radial basis function networks, neural networks, and locally weighted regression were applied in this problem domain. The neural networks with backpropagation give the best results in terms of performance on the original dataset with average values for the target attribute. The experiments suggest that finding the right feature subset is more important than finding the right machine learning scheme for this problem domain. The uncertainty handling capability of the machine learning schemes were also tested. It was found that the uncertainty handling should be incorporated in the feature selection process as well.