Support Vector Machine-based Aproach for Multi-labelers Problems

We propose a first approach to quantify the panelist’s labeling generalizing a soft-margin support vector machine classifier to multi-labeler analysis. Our approach consists of formulating a quadratic optimization problem instead of using a heuristic search algorithm. We determine penalty factors for each panelist by incorporating a linear combination in the primal formulation. Solution is obtained on a dual formulation using quadratic programming. For experiments, the well-known Iris with multiple simulated artificial labels and a multi-label speech database are employed. Obtained penalty factors are compared with both standard supervised and non-supervised measurements. Promising results show that proposed method is able to asses the concordance among panelists considering the structure of data.