Clustering and combinatorial optimization in recursive supervised learning

摘要: The use of combinations weak learners to learn a dataset has been shown be better than the single strong learner. In fact, idea is so successful that boosting, an algorithm combining several for supervised learning, considered best off shelf classifier. However, some problems still exist, including determining optimal number and over fitting data. earlier work, we developed RPHP which solves both these by using combination global search, learning pattern distribution. this chapter, revise search component replacing it with cluster based combinatorial optimization. Patterns are clustered according output space problem, i.e., natural clusters formed on patterns belonging each class. A optimization problem therefore created, solved evolutionary algorithms. algorithms identify “easy” “difficult” in system. removal easy then gives way focused more complicated patterns. becomes recursively simpler. Over overcome set validation along distributor. An also proposed distributor determine recursions hence problem. Empirical studies show generally good performance when compared other state art methods.