Selecting machine learning models for metallic nanoparticles

Abstract

The outcome of machine learning is influenced by the features used to describe the data, and various metrics are used to measure model performance. In this study we use five different feature sets to describe the same 4000 gold nanoparticles, and 14 different machine learning methods to compare a total of 70 high scoring models. We then use classification and regression to show which meta-features of data sets or machine learning algorithms are important when making a selection. We find that number of features, and those that are strongly correlated, determine the class of model that should be used, but overall quality is almost entirely determined by the cross-validation score, regardless of the sophistication of the algorithm.