Function Approximation for Adaptive Learning of Label Distributions

Rather than traditional pattern and signal analysis problems, label distribution learning aims to predicate the probabilistic labels of appended instances. In this work, a novel approach to LDL is proposed for adaptive learning of label distributions of instances, which adopts function approximation to label predication. As a consequence, the predication of probabilistic labels of instances are able to be accomplished with self-taught optimization of approximate factorizations, and stable calculation efficiency can be achieved. Experiments on diverse artificial data sets demonstrate the proposed method is able to give stable complexities on different data sets while comparable performance is obtained, and insensitive to sizes of data sets.