Knowledge discovery by fusion of information.

摘要

Information fusion aims to develop intelligent approaches of integrating information from different complementary sources, so that a more comprehensive basis is obtained for data analysis and knowledge discovery. These approaches are particularly interesting in the areas of geo-informatics and text mining in bioinformatics, where huge size of high-dimensional data from multiply correlated sources are available while their inherent relationships have not been well understood. In the first part of the thesis, we study the data-driven approaches to improve aerosol optical thickness (AOT) retrieval performance from satellite-based and ground-based observations. We explore the statistical models which complement deterministic retrieval algorithms to reduce computational costs on huge size of remote sensing data. The experiments showed that, given a small fractions of deterministic retrievals for training, statistical models can significantly speed up AOT retrievals while introducing only a slight accuracy overhead. Next, in order to construct an accurate and understandable AOT predictor, we combine spatially and temporally co-located data from multi-sources into a uniform dataset. Artificial neural networks (ANNs) are applied to derive optimal regression models. The results suggest ANN models achieve overall accuracy superior or comparable to deterministic AOT retrievals. The decision tree analysis reveals ANN predictions effectively enhance deterministic AOT retrievals in some surface or climate conditions. The second part of the thesis addresses the problems of identifying biomedical publications with desired experimental evidence from MEDLINE, a major biomedical repository collecting millions of domain papers from different journals and conference proceedings. The learning task is challenged by richness of biomedical terminology sources, diversity of experimental evidence expressions and a small list of labeled examples for training. We propose a novel substring construction algorithm which derives attributes from semantically-related terms with shared stems or morphemes. With five post-translational modification (PTM) test datasets, curators confirm the selected substrings significantly improve classification performance. Finally, we summarize our work and propose future research directions. Specifically, we describe a framework which exploits online text to explain some aerosol results of satellite image analysis.