Kernelized Extension for Multiple-Expert-Learning Classifiers Using Support Vector Machines

The primary goal for supervised machine learning techniques is to make accurate predictions or classifications based on prior knowledge about the relationships between the input features and their corresponding label assignment. The latter is often difficult to be determine, especially when the ground truth is not unique but instead, a combined decision from a group of experts is given. One way to tackle this challenge is by using Multiple Expert Learning (MEL)-driven classifiers. In this regard, approaches based on Soft-margin support vector machines (SM-SVM) including penalty factors for each labeler have shown to be a suitable alternative as they can learn from experts holding different levels of trustworthiness. Nonetheless, dealing with nonlinear (complex structure) and hardly separable data is still an open issue. To address these problems, this work presents mathematical developments on SVM for MEL scenarios in a fully matrix formulation. As a remarkable contribution, we introduce a kernel extension of the SM-SVM problem for MEL scenarios together with a complete solution following from a quadratic programming approach.