Scalable Learning with Ensemble Techniques and Parallel Computing

描述（由申请人提供）：开展基础和应用生物医学研究的能力越来越依赖于新兴技术产生的数据。这些数据具有前所未有的详细程度和数量。因此，研究人员依赖于计算和计算工具来可视化、分析、建模和解释这些庞大而复杂的数据集。疾病检测、诊断、治疗和预防工具是许多生物医学研究项目的共同目标，如果不是全部的话。用于类预测和类发现的可靠的分析和统计理论和方法为构建这些工具奠定了基础，其中分类（监督学习）和聚类（无监督学习）的机器学习技术至关重要。我们的目标是使用集成机器学习技术和并行计算技术生产用于分析和解释大型数据集的软件。集成技术是机器学习理论和方法的最新进展，它极大地提高了数据集分析和解释的准确性和稳定性。在不同的实例或特征子集上训练的主要机器学习器（分类器或聚类器）委员会的结果通过投票等技术进行组合。分类器集成（如boosting、bagging和random forest）的高预测精度在统计学和机器学习社区中引起了很大的兴奋。最近的研究将集成方法扩展到类信息不可用的聚类，在准确性和稳定性方面也产生了卓越的性能。理论上，大多数集成技术本质上是并行的。但是，现有的实现通常是串行的，并且假设数据集驻留在内存中。因此，当前的软件不能扩展到今天生物医学研究中产生的大数据集。我们建议采用两种方法将集成技术扩展到大型数据集：数据分区方法和并行计算。第一阶段的重点将是使用并行架构构建可扩展分类器集成的原型。我们打算：建立并行计算基础设施；生成初步的架构和软件设计；研究使用数据分区策略的各种集成生成方案；并在初步设计的基础上实现可扩展套袋和随机森林。第二阶段的重点将是完成软件体系结构，并在该框架内实现可伸缩的分类器集成和可伸缩的集群集成。我们打算：完成分类器集成的研究和开发；将分类框架扩展到聚类集成；研究和开发具有不同生成机制和组合策略的建筑集成的统一接口；并通过仿真数据和实际数据对软件的有效性进行了评价。公共卫生相关性：许多（如果不是全部的话）生物医学研究项目的共同目标是开发疾病检测、诊断、治疗和预防工具。这些程序通常依赖于具有前所未有的细节和数量的新类型数据。我们的目标是使用集成机器学习技术和并行计算技术生产用于分析和解释大型数据集的软件，使依赖计算工具的研究人员能够可视化，分析，建模和解释这些大型和复杂的数据集。

DESCRIPTION (provided by applicant): The ability to conduct basic and applied biomedical research is becoming increasingly dependent on data produced by new and emerging technologies. This data has an unprecedented amount of detail and volume. Researchers are therefore dependent on computing and computational tools to be able to visualize, analyze, model, and interpret these large and complex sets of data. Tools for disease detection, diagnosis, treatment, and prevention are common goals of many, if not all, biomedical research programs. Sound analytical and statistical theory and methodology for class pre- diction and class discovery lay the foundation for building these tools, of which the machine learning techniques of classification (supervised learning) and clustering (unsupervised learning) are crucial. Our goal is to produce software for analysis and interpretation of large data sets using ensemble machine learning techniques and parallel computing technologies. Ensemble techniques are recent advances in machine learning theory and methodology leading to great improvements in accuracy and stability in data set analysis and interpretation. The results from a committee of primary machine learners (classifiers or clusterers) that have been trained on different instance or feature subsets are combined through techniques such as voting. The high prediction accuracy of classifier ensembles (such as boosting, bagging, and random forests) has generated much excitement in the statistics and machine learning communities. Recent research extends the ensemble methodology to clustering, where class information is unavailable, also yielding superior performance in terms of accuracy and stability. In theory, most ensemble techniques are inherently parallel. However, existing implementations are generally serial and assume the data set is memory resident. Therefore current software will not scale to the large data sets produced in today's biomedical research. We propose to take two approaches to scale ensemble techniques to large data sets: data partitioning approaches and parallel computing. The focus of Phase I will be to prototype scalable classifier ensembles using parallel architectures. We intend to: establish the parallel computing infrastructures; produce a preliminary architecture and software design; investigate a wide range of ensemble generation schemes using data partitioning strategies; and implement scalable bagging and random forests based on the preliminary design. The focus of Phase II will be to complete the software architecture and implement the scalable classifier ensembles and scalable clusterer ensembles within this framework. We intend to: complete research and development of classifier ensembles; extend the classification framework to clusterer ensembles; research and develop a unified interface for building ensembles with differing generation mechanisms and combination strategies; and evaluate the effectiveness of the software on simulated and real data. PUBLIC HEALTH RELEVANCE: The common goals to many, if not all, biomedical research programs are the development of tools for disease detection, diagnosis, treatment, and prevention. These programs often rely on new types of data that have an unprecedented amount of detail and volume. Our goal is to produce software for the analysis and interpretation of large data sets using ensemble machine learning techniques and parallel computing technologies to enable researchers who are dependent on computational tools to have the ability to visualize, analyze, model, and interpret these large and complex sets of data.