In my dissertation, I developed a novel approach to learning from data based on a formulation of the learning task in terms of manipulation of answers to a succession of statistical queries against the data source. This allows us to reduce the problem of learning from distributed data into the problem of answering statistical queries from distributed data. I identified the statistical queries needed by a broad class of machine learning algorithms and developed provably sound approaches to answering such queries from distributed data. This approach offers a general strategy for transforming traditional algorithms for learning from data into algorithms for learning from distributed data.

I developed approaches to decomposing a statistical query posed by the learner into sub-queries according to the distributed data sources, finding the best execution plan, executing this plan and composing the individual answers obtained from the distributed data sources into a final answer to the initial query.  My dissertation illustrates the application of this strategy to devise several algorithms (Naïve Bayes, Decision Trees, Perceptron, Support Vector Machines and k-NN) for induction of classifiers from horizontally and vertically fragmented distributed data. The resulting algorithms are provably exact in that the classifiers produced by them are identical to those obtained by the corresponding algorithms in the centralized setting (i.e., when all of the data is available in a central location). This ensures that the entire body of theoretical (e.g., sample complexity, error bounds) and empirical results obtained in the centralized setting carry over to the distributed setting. The proposed algorithms compare favorably to their centralized counterparts in terms of time and communication complexity in many scenarios that are common in practice.

I extended this approach to learning from distributed data to learning from semantically heterogeneous distributed data, where semantic differences between data sources and the user are inevitable. To deal with semantic heterogeneity, I introduced ontology extended data sources (with explicit representation of the ontology associated with each data source).  In my dissertation, I showed how the user-specified interoperation constraints between user ontology and data source specific ontologies can be used to define mappings and conversion functions needed to answer statistical queries from semantically heterogeneous data sources from a user perspective, in the important special case where ontologies take the form of attribute value hierarchies.  The resulting approach to learning classifiers from semantically heterogeneous data allows flexible data integration on demand during the process of answering statistical queries needed by a learning task. This approach is especially well suited for scientific applications that call for exploratory analysis of semantically heterogeneous data sources from different perspectives.

During the next few years, I plan to explore:

1) Extensions of my approach to learning from semantically heterogeneous data in settings where ontologies take more general forms than the attribute value taxonomies treated in detail in my dissertation.

2) Extensions of my approach to learning classifiers from semantically heterogeneous data to learning complex relationships from semantically heterogeneous multi-relational data.

3) Provably approximate algorithms for learning classifiers from distributed semantically heterogeneous data in settings where the computational requirements of provably exact algorithms are prohibitive.

4) Sufficient statistics based approaches to learning from data streams where it is infeasible to store the data.

5) Application of the algorithms for learning classifiers from semantically heterogeneous, distributed data to knowledge acquisition tasks that arise in computational biology (e.g., protein function/structure prediction, protein-protein interaction sites prediction).

Research in the area of knowledge acquisition from semantically heterogeneous data sources is still in its infancy. It presents research challenges that span multiple areas of computer science including machine learning, knowledge representation and inference, databases, information integration, and statistical computing. My early contributions to this area put me in an excellent position to tackle some of these challenges.