Statistical learning: methodology and efficient algorithms

我的研究目标是开发高效的监督和非监督统计学习工具，用于分析药物发现和调查抽样的数据。 我在药物发现方面的研究主要集中在发明有监督的学习方法，以捕捉药物发现数据的复杂结构，并将稀有活性化合物(稀有靶标问题)识别为潜在的候选药物。药物发现中的统计学习寻求建立一个模型，该模型使用表征分子结构的描述符(解释变量)来预测生物活性(响应变量)。然而，药物发现数据的特点使得很难对分子描述符和生物活性之间的关系进行建模。为了解决上述问题，提出了一种基于混合模型的监督学习方法，该方法考虑了描述子的测量误差。这包括确定混合模型中组分数量的模型选择、测量误差建模和开发高效的计算机算法。 我对抽样调查的研究将集中在发明无监督统计学习方法来聚类具有有序分类变量和大量缺失值的抽样调查数据。我将重点介绍当测量的变量可能是带有缺失值的分类排序时的观察结果。在贝叶斯框架下，提出了一种基于模型的聚类算法。

The objective of my research is to develop efficient supervised and unsupervised statistical learning tools for the analysis of data from drug discovery and survey sampling. My research on drug discovery focuses on inventing supervised learning methods to capture the complex structure of drug discovery data and identify rare active compounds (rare target problem) as potential drug candidates. Statistical learning in drug discovery seeks to build a model that uses descriptors characterizing molecular structure (explanatory variables) to predict biological activity (response variable). However, the characteristics of drug discovery data can make it difficult to model the relationship between molecular descriptors and biological activity. A supervised learning method based on the mixture model with the consideration of the measurement error of descriptors will be developed to deal with above challenges. This involves model selection determining the number of components in the mixture models, measurement error modeling and development of efficient computer algorithms. My research on the sample survey will focus on inventing unsupervised statistical learning methods to cluster sample survey data with ordered categorical variables and a large amount of missing values. I will focus on clustering observations when variables measured may be ordered categorical with missing values. a model-based clustering algorithm will be developed in Bayesian framework.