Simulator for multi-labeler scenario

Iván Darío Gustin and Mauricio Bolañoz Ledezma, Universidad de Nariño, San Juan de Pasto - Colombia 2015

Typically, supervised pattern recognition systems are trained by using prior knowledge-expressed by labels- given by a single expert labeler or annotator. Nonetheless, for some applications is suitable to consider a set of labelers rather than only one. For instance, a set of specialists diagnosing a patient’s pathology or a teacher team assessing the academic performance of a student. Despite it is necessary, considering multiple labelers makes the classification problem difficult since there is no a clearly, identified ground truth. 

To classify data within this scenarios, multi-labeler strategies have been proposed, which should be able to both compensating the influence of wrong labels regarding the assumed ground truth, as well as identifying the good and bad labelers. In this connection, support-vector-machines- (SVM) based approaches have shown to be a suitable alternative. Most of the currently available methods make strong assumptions on the resultant labeling vector, introducing then naturally noise over the classification task. 

In this work, we present a novel approach for data classification within a multi- labeler approach. The classification is done by making a mixture of classifiers  trained by using each labeler. Therefore, there is no assumptions made on the estimation of ground truth. Our method just naturally obtains an adjusted objective function defining an improved boundary decision.

Figure 1. Weighted Mixture Classifiers



        The 21st IberoAmerican Congress on Pattern Recognition, CIARP-2016            
        I. D. Gustin, D. E. Imbajoa-Ruiz, M. Bolaños-Ledezma, F. Arciniegas-Mejia, F.A. Guasmayan-Guasmayan, M. J. Bravo-Montenegro, A.E.Castro-Ospina, D.H. Peluffo-Ordoñez

        International Conference on Information Systems and Computer Science, INCISCOS 2016
        I. D. Gustin, D. E. Imbajoa-Ruiz, M. Bolaños-Ledezma, F. Arciniegas-Mejia, Dario F. Fajardo F, F.A. Guasmayan-Guasmayan, M. J. Bravo-Montenegro, D.H. Peluffo-Ordoñez


Figure 2. Multi-labeler Simulator

We designed a simulator pattern recognition within a multi-expert scenario using the mixing method based on the weighted average approach (see Figure 2), you can download the script at the following link.

Video Tutorial

User Manual

The user manual can be downloaded from the following link:

About US

 Ivan D. Gustin Mauricio Bolaños Ledezma
 Electronic Engineer  Electronic Engineer