Active Learning of Qualitative Models with Pad´ e
Tadej Jane, Martin Mo, Ivan Bratko · 2010
This paper presents a new approach to learning qualitative models, namely active learning of qualitative models. We combine a traditional approach of learning qualitative models from numerical data with the active learning paradigm. In general, active learning is useful when labeling learning examples is expensive or difficult. Selecting a representative subset of learning examples is thus extremely important in order to learn a good model. While learning qualitative models by approximating partial derivatives of output variable w.r.t. the selected input variable, we can use the ceteris paribus assumption to choose the next learning example for which we demand a label. In other words, the next most useful example is the one that changes the value in the direction of differentiation, other variables being equal. We propose a new method, = ;