Abstract
A number of machine learning algorithms have been proposed and deliver superior empirical performances. However, the understanding of their mechanisms is hampered by the black-box nature of the algorithms. In this dissertation, we approach the problem from several different lens. One is visualization, with a data-driven geometry following kernel — the rpf-kernel, which can extract complex and highly nonlinear patterns in the data beyond the usual principal component analysis. The second is the diagnosis perspective. Specifically, we perform a diagnostic analysis to data points under a given algorithm and hope to use this as a proxy to understand the algorithm. Random Forests classification is used as an example algorithm for our study. We borrow two metrics, leverage and influence, from statistics regression to measure the importance of data points, while extending their definition to a small neighborhood of data points. Variable importance is another aspect we study. It is of major significance in the practice of statistical analysis and model interpretation. However, current methods do not consider the correlation between variables, and the importance of a given variable tends to be “masked” by correlated variables. We proposed a de-correlation based approach to solve this problem and obtained a more reasonable importance score for variables in the model. Also studied is a related issue of fairness — whether the algorithm delivers a response that is fair in terms of some given metric, for example the gender of the associated subjects. K-means clustering is studied, and a computationally efficient post algorithm adjustment is proposed. Experiments show that the proposed method is effective in improving the fairness while maintaining the clustering performance.