Many applications of data-driven knowledge discovery processes call for the exploration of data from multiple points of view that reflect different ontological commitments on the part of the learner. Of particular interest in this context are algorithms for learning classifiers from ontologies and data. Against this background, my dissertation research is aimed at the design and analysis of algorithms for construction of robust, compact, accurate and ontology aware classifiers. We have precisely formulated the problem of learning pattern classifiers from attribute value taxonomies (AVT) and partially specified data. We have designed and implemented efficient and theoretically well-founded AVT-based classifier learners. Based on a general strategy of hypothesis refinement to search in a generalized hypothesis space, our AVT-guided learning algorithm adopts a general learning framework that takes into account the tradeoff between the complexity and the accuracy of the predictive models, which enables us to learn a classifier that is both compact and accurate. We have also extended our approach to learning compact and accurate classifier from semantically heterogeneous data sources. We presented a principled way to reduce the problem of learning from semantically heterogeneous data to the problem of learning from distributed partially specified data by reconciling semantic heterogeneity using AVT mappings, and we described a sufficient statistics based solution.