University of Houston

There has been a steady grow of interest in the utilization of machine learning algorithms to solve real problems in a wide variety of fields. The successful use of learning algorithms outside the boundaries of research (e.g., industry, commerce, government) is conditioned on the appropriate selection of a suitable predictive model (or combinations of models). There is a need for automated tools that can provide systematic user guidance on model selection. The problem of model selection can be tackled through different methods (e.g. empirical, theoretical, meta-learning based, etc). One major missing ingredient in these approaches is a systematic study of the locality concept. The concept of locality has gained importance due to the complex nature of real problems; we plan to study such locality using a divide-and-conquer approach. In general, we advocate tackling the model selection problem from a local perspective. This research proposes three different local perspectives to model selection: i) a regularization framework that includes a term for locality; ii) a model selection algorithm for hierarchical classification that is capable of modifying the hierarchy to improve performance; iii) a meta-learning based local model selection algorithm that exploits existing knowledge about the relationship between strengths and limitations of learning algorithms and local data properties.