Proc. SPIE. 9662, Photonics Applications in Astronomy, Communications, Industry, and High-Energy Physics Experiments 2015
KEYWORDS: Data modeling, Binary data, Performance modeling, Machine learning, Statistical modeling, Data acquisition, Computer science, Computing systems, Systems modeling, Geographic information systems
Boosting algorithms constitute one of the essential tools in modern machine-learning, one of its primary applications being the improvement of classifier accuracy in supervised learning. Most widespread realization of boosting, known as AdaBoost, is based upon the concept of building a complex predictive model out of a group of simple base models. We present an approach for local assessment of base model accuracy and their improved weighting that captures inhomogeneity present in real-life datasets, in particular in those that contain geographic information. Conducted experiments show improvement in classification accuracy and F-scores of the modified algorithm, however more experimentation is required to confirm the exact scope of these improvements.