Collaborative Research "Tracking Statistics and Inference for Indirect Measurements"

分类和聚类是发现有用模式的两种基本数据挖掘工具。鉴于通常不存在完美的分类或聚类程序，在从分类结果导出的任何后续推断中校正或解释分类错误是至关重要的。 该提案中的研究正在开发推理程序，其中包含与分类率相关的错误，从而提高基于分类和聚类机制的决策过程的鲁棒性。 一个特别的例子是在过程控制应用中开发跟踪统计，以纠正缺陷分类中的错误。 另一个例子是在没有通常假设金本位存在的情况下，发展错误分类率估计。 这项研究通过消除对数据质量水平的严格要求，使数据挖掘和知识发现技术得到更广泛的应用。 许多决策过程使用的输入是根据相似性对主题进行分组的统计分析的结果。 分类和聚类技术是两种重要的分组方法。决策的有效性取决于分组结果的准确性。 如今，分类和聚类算法使用的大型数据库通常数据质量较低，因此在分类和聚类结果中引入了偏差。 本研究的目标是开发推理方法，调整固有的噪声在分类和聚类算法的输出，从而提高后续决策的准确性。这项研究的结果应该有利于许多领域的应用，其中包括微阵列基因表达数据的分析，机器学习，信息检索，风险分析，计算机辅助诊断和模式识别。

Classification and clustering are two fundamental data mining tools for discovering useful patterns. Given that there is generally no perfect classification or clustering procedures, it is crucial to correct or account for the classification error in any subsequent inference which is derived from the classification outcomes. The research in this proposal is developing inference procedures that incorporate the error associated with classification rates, and consequently is improving the robustness of decision-making processes that are based on classification and clustering mechanisms. A particular example is the development of tracking statistics in process control applications that correct for errors in defect classifications. Another example is the development of misclassification rate estimates without the usual assumption that a gold standard exists. The research is enabling a wider use of data mining and knowledge discovery techniques by removing stringent requirements on data quality levels. Many decision-making processes use inputs that are the result of statistical analyses of grouping subjects according to their similarities. Classification and clustering techniques are two important such grouping methods. The validity of the decision-making rests on the accuracy of the grouping outcomes. Nowadays, classification and clustering algorithms make use of large databases that are often low in data quality, and consequently introduce biases in the classification and clustering outcomes. The goal of this research is to develop inference methodologies that adjust for inherent noise in the outputs of classification and clustering algorithms, and thereby improve the accuracy of subsequent decision-making. The results of this research should benefit many areas of applications, which include the analysis of micro-array gene expression data, machine learning, information retrieval, risk analysis, computer-aided diagnostics, and pattern recognition.