Justin Valentine 501 Presentation

Title: Sampling Methods for Imbalanced Classification

Abstract: Classification predictive modeling poses unique challenges when the number of observations in each class is imbalanced. Many of the most common learning algorithms were designed for use on balanced datasets which can sometimes result in poor performance yet misleadingly optimistic results. This can be addressed by transforming the dataset or by changing how the algorithm learns. Various types of data sampling techniques have been developed to balance the dataset with artificial values that closely resemble the data. An alternative to this is cost-sensitive learning. Methodologies for these techniques will be explained and showcased using real world data, and model performance evaluated using performance metrics more suited for imbalanced classification.