Introducing fuzzy decision stumps in boosting through the notion of neighborhood

This paper presents a novel approach to the weak classifier selection based on the GentleBoost framework. We include explicitly the notion of neighborhood in one of the most common weak learner in boosting, the decision stumps. The availability of neighboring points adds a new parameter to the decision stump, the feature set (i.e. neighborhood), and turns the single branch selection of the decision stump into a fuzzy decision that weights the contribution of each branch using a neighborhood-based confidence measure. The confidence measure of the fuzzy stumps use neighboring samples to increase the robustness to local data perturbations. The appropriate definition of the neighborhood in the dataset allows the application of the fuzzy stumps framework in a wide range of problems. In this paper we address two types of scenarios to show their advantages: i) time-based neighborhoods and ii) space-based neighborhoods. In both scenarios we evaluate experimentally the properties of the fuzzy stumps, considering computer generated datasets and real classification problems, such as human activity recognition and object detection.