M.S. in Engineering
Electrical & Computer Engineering
Henry M. Rowan College of Engineering
Electrical engineering--Research; Learning--Mathematical models
Electrical and Computer Engineering
Many pattern classification problems require a solution that needs to be incrementally updated over a period of time. Incremental learning problems are often complicated by the appearance of new concept classes and unbalanced cardinality in training data. The purpose of this research is to develop an algorithm capable of incrementally learning from severely unbalanced data. This work introduces three novel ensemble based algorithms derived from the incremental learning algorithm, Learn++. Learn++.NC is designed specifically for incrementally learning New Classes through dynamically adjusting the combination weights of the classifiers' decisions. Learn++.UD handles Unbalanced Data through class-conditional voting weights that are proportional to the cardinality differences among training datasets. Finally, we introduce the Boosted Ensemble Algorithm Strategically Trained (BEAST) for incremental learning of unbalanced data. BEAST combines Learn++.NC and Learn++.UD with additional strategies that compensate for unbalanced data arising from cardinality differences among concept classes. These three algorithms are investigated both analytically and empirically through a series of simulations. The simulation results are presented, compared and discussed. While Learn++.NC and Learn++.UD perform well on the specific problems they were designed for, BEAST provides a strong and more robust performance on a much broader spectrum of complex incremental learning and unbalanced data problems.
Muhlbaier, Michael David, "Boosted ensemble algorithm strategically trained for the incremental learning of unbalanced data" (2006). Theses and Dissertations. 915.