Flexible Statistical Modeling

DMS-9803645Hastie在过去的10-15年里，在应用回归和分类领域有了重大发展。大部分推动力最初来自统计领域以外的领域，如计算机科学、机器学习和神经网络。这些学科带来了许多新的想法，许多新的和令人兴奋的模型，如神经网络，以及许多有趣的应用领域。随着尘埃落定，我们发现，这些新想法最好是在统计框架内综合，并与传统的线性和非线性模型相提并论。这项研究计划的一个关键项目是一本研究专著，其暂定标题是：统计学习的要素(与Jerome Friedman和Rob Tibishani合著)。这本书开发了一个框架，用于从统计的角度描述和理解新的回归和分类技术，并将它们与现有方法相结合。我们在经典的经过充分测试的线性和参数模型与更具新奇和适应性的技术之间取得了自然的平衡，这些技术适用于数据丰富的场景。数据分析和建模中的许多重要问题都集中在预测上：计算机辅助疾病诊断(例如读取数字乳房X光照片)、心脏病风险评估、自动读取手写数字(例如信封上的邮政编码)、语音识别等等。这个研究项目有两个方面。第一本是一本专著，从许多不同的贡献中合成了一系列经过良好测试的技术，并从统计学的角度对它们进行了解释。第二个武器是开发一些新的预测技术。所有这些新方法都利用了我们可用的快速计算设施，并使我们能够开发出十年前不可能实现的预测方法。

DMS-9803645HastieThere have been significant developments in the areas of applied regression and classification over the past 10-15 years. Much of the impetus originally came from outside of the field of statistics, from areas such as computer science, machine learning and neural networks. These disciplines have brought many fresh ideas to the table, a host of new and exciting models such as neural networks, as well as many interesting areas of application. As the dust settles, we find that these new ideas are best synthesized within a statistical framework, and have a natural place alongside traditional linear and nonlinear models. A key item in this research program is a research monograph with working title: THE ELEMENTS OF STATISTICAL LEARNING (with Jerome Friedman and Rob Tibshirani). This book develops a framework for describing and understanding the new regression and classification techniques from a statistical point of view, and for synthesizing them with existing methods. We strike a natural balance between the classical well tested linear and parametric models, and the more exotic and adaptive techniques appropriate in data rich scenarios. The research program includes the development of some new techniques for multiclass classification, each of which expand on existing techniques in novel ways.Many important problems in data analysis and modeling focus on prediction: computer assisted diagnosis of disease (e.g. reading digital mammograms), heart disease risk assessment, automatic reading of handwritten digits (e.g. zip-codes on envelopes), speech recognition, to name a few. This research program has two arms. The first is a monograph that synthesizes from the many varied contributions a collection of well-tested techniques, and explains them from a statistical point of view. The second arm is to develop some new techniques for prediction. All these new methods exploit the rapid computing facilities we have available, and allow us to develop methods for prediction that would have been infeasible ten years ago.