Flexible Classification and Regression

点击翻译按钮获取中文摘要

The research aims to combine statistical and computational considerations in designing new and useful predictive modeling tools and algorithms. Specifically, the research involves the development of: a) new statistically motivated multi-class boosting algorithms, based on a family of multi-class loss functions and forward stagewise additive modeling; b) a family of (loss, penalty) pairs that give piecewise linear solution paths, and yield modeling tools for both regression and classification which are robust, adaptable and efficient; c) a general theory and efficient algorithms for solving an L1 regularized problem in infinite dimensional predictor space.With the advent of modern technologies, the needs for predictive modeling tools have been increasing rapidly. Consequently, many new ideas and methods have been finding their way into the statistical community in recent years. These are mainly related to the design and analysis of useful techniques for modeling of high dimensional, noisy data, and these techniques are now being applied to bioinformatics, high energy physics, speech recognition, text mining, and a wide range of other important practical problems. This research aims to push these developments forward along the line of regularization in predictive modeling, and is expected to have broader impacts on the practice and education in the domains of statistics, machine learning and data mining.