They propose new methods to improve the automatic learning of the machines
7 January 2013
Mikel Galar Idoate, computer engineer and doctor by the Public University of Navarra, has proposed new methods to obtain better results in the automatic classification. At present it works in a project of investigation on autenticación biométrica by means of footprints dactilares. The automatic learning of the machines, that is to say, treat that the machines learn of automatic way from situations already known previously, has been the field of investigation of Mikel Galar Idoate, that has read his thesis doctoral in the Public University of Navarra.
In concrete, the investigation of Mikel Galar has centred in the problems of classification. “We imagine that we study a type of cancer —explains—: we extract the data of some fabrics and define to which class can belong. By means of the automatic classification, the machine treats to assign to each fabric one of the classes predefinidas: cancerígeno, benign or malignant. What try is that the machine learn to classify new examples, having as basic examples of the problem that already know”.
The advantage that have the automatic methods is that the analysis of the data lacks the inherent subjectivity to the human being. “Besides, with an automatic method the capacity of analysis and the volume of data with which can work always are a lot greater that the ones of a person”. They exist crowd of problems of classification in which it can use this type of technicians: banking, medicine, bioinformática or, of way more specific, detection of delinquent, classification of footprints dactilares, diagnosis of cancer, detection of spam in ecourierss, etc.
One of the most used technicians in the last years to face the problem of the automatic classification consists in using ensembles or groups of sorters. “Of similar way to what do the human beings, that consult to a series of experts before taking an important decision, with the use of a group of sorters tries classify examples of the same problem, combine the answers or exits and obtain like this better decisions of which would obtain if we used an only sorter”.
The work of Mikel Galar has centred in three of the areas where the use of ensembles has been beneficial: problems of classification with multiple classes, the problem of the no swung classes and the problem of the difficult classes. “They are key problems in the automatic learning. In the development of the thesis doctoral, have analysed each area, his fortresses and weaknesses, and have proposed new methods that have obtained better results to face the existent problems until the moment”. Like result of the thesis have arisen five articles in magazines of international recognition, as well as several communications in international conferences.
The thesis, titled ‘Ensembles of classifiers for multiclass classification problems: one-vs-one, imbalanced dates-sets and difficult classes', has been directed by the doctors Edurne Barrenechea (UPNA), Alberto Fernández (University of Jaén) and Francisco Herrera (University of Granada) and has obtained the qualification cum laude with international quotation.