9310413 Pazzani The objective of the proposed research are to investigate the utility of algorithms that learn probabilistic relational concepts from noisy, relational data, concentrating on algorithms that perform classification using evidence combination. Of particular interest is the potential improvement in classification accuracy that can be obtained if the reliability of rules is taken into account. An additional objective is to extend noise-tolerant learning algorithms to take advantage of incomplete or uncertain knowledge of the concept to be learned. Such knowledge may be expressed as a partially incorrect or incomplete domain theory. These algorithms will be tested on a number of data sets from the UCI Repository of Machine Learning data sets and on three new relational data sets involving spinal injury identification, mineral spectra identification, and oil spill analysis. In addition, they will be tested on a number of synthetic domains, where variation of parameters will allow insight into the general behavior of these algorithms. The system HYDRA (Ali & Pazzani, 1993), which has already shown some promise on a limited number of data sets will be significantly extended for these investigations.
