Data fusion in scientific data mining.

摘要

Data fusion involves multi-sources or multi-presentations of a single source to perform inferences which are more comprehensive and accurate than those of any single method. Thus, data fusion makes it possible to create a synergistic process in which the consolidation of individual data creates a combined resource with a productive value greater than the sum of its parts.;While considerable research has been done on data fusion in the past, most of them performed in the field of multi-sensor fusion. There has been relatively less work conducted in a data mining context. The goal of this dissertation is to develop a data fusion framework for predictive modeling, especially to the Quantitative Structure-Activity Relationship (QSAR) problems, which includes a function-oriented model, general architecture paradigms and corresponding learning algorithms. Furthermore, kernel methods, e.g. kernel partial least squares (K-PLS) ensemble with bagging and boosting is introduced as one of important decision level fusion methods. This approach can be applied to applications with multiple data sources available to obtain information of greater quality. In addition to the predefined three fusion levels, the kernel fusion method is further developed based on the properties of kernel in the feature space to take advantages of multiple physically different feature sets in order to build more accurate and robust predictive models. With Hessian-free and self-correction properties, the BFGS quasi-Newton method is employed for parameter tuning of kernel fusion.;In addition to the regression algorithm applied in data fusion scheme introduced above, we also are trying to extend K-PLS to the classification, especially when the class distribution is highly skewed or changes dramatically over time. In this case, a probabilistic classifier with the capability to deal with high dimensional discrimination is desirable. To achieve this goal, a new kernel orthonormalized PLS logistic regression (KOPLS-LR) and the corresponding ROC based adaptive threshold (ROC-BAT) approach are proposed. KOPLS-LR inherits the advantages of K-PLS and logistic regression, while ROC-BAT provides an effective solution to predict observations in which the distribution dramatically changes over time.;During the research process, a web-based modeling system has been designed and realized which integrates various learning methods, e.g. PLS, K-PLS and support vector machines (SVM). Model selection (parameter tuning) and performance estimation functionalities are also integrated in this online predictive modeling system. This predictive tool is open to the public and can be accessed at: http://reccr.chem.rpi.edu/Software/modeling/index.html