Fusion of Continuous‐Valued Outputs
Ludmila Ilieva Kuncheva · 2004
Chapter 5 presents methods for fusing continuous-valued classifier outputs (degrees of support for the classes) based on the notion of decision profile. We explain how classifier outputs can be transformed into probabilities and proceed to introduce class-conscious and class-indifferent combiners. Simple combination schemes (called also combiners, combination rules or aggregation rules) are listed including mean (average or sum), product, median, minimum, maximum and a class of generalized mean combiners. Next we introduce trainable combiners including weighted average, fuzzy integral, decision templates and a Dempster–Shafer combiner. Finally we present several theoretical models behind simple combiners coming from various perspectives: conditional independence, Bayesian and supra Bayesian approaches, minimization of Kullback–Leibler divergence and consensus theory.