"Several operations of recognition and prediction are performed nowadays, many without even people consciousness. Predictive learning has traditionally consisted in constructing rules which discriminate positive from negative, or malign from benign cases depending on the scenario in analysis. Models are constructed by gathering knowledge from data. Data describe the problem through different features, attributes or criteria thereby constituting the feature space. The insight gained will make possible the estimation of a mapping from the feature space into a finite class space. Depending on the cardinality of the finite class space we are left with binary (e.g., positive and negative) or multiclass classification problems. In more complex situations, one has to deal with data where, the presence or absence of a «natural» order among classes, will separate ordinal from nominal problems.
Retrieving information in a way that we can interpret different criteria on data has been playing major roles in the academy and industry. Specially in scenarios where data contains an order relation not only on the classes labels but also on the data itself. Learning models for these settings are referred to as ordinal data problem. The credit scoring problem is an example of that. In this problem, one evaluates how unlikely a client will default with his payments. Client profiles are evaluated, being their results expressed in terms of an ordinal score scale (Excellent > Good > Fair > Poor). Intelligent systems have then to take into consideration different criteria such as payment history, mortgages, wages among others in order to accomplish their outcome.
Contributions of this work are three fold. Firstly, we have shown that existing measures for evaluating ordinal classification models suffer from a number of important shortcomings. For this reason, we proposed an alternative measure defined directly in the confusion matrix. An error coefficient appropriate for ordinal data was therefore designed such that it captures how much the result diverges from the ideal prediction and how «inconsistent» the classifier is in regard to the relative order of the classes.
Secondly, we have identified that despite the myriad of schemes for multi-class classification with SVM, little work has been done for the case where the classes are ordered. Hence, a new SVM methodology was proposed based on the unimodal paradigm with the All-at-Once approach for the ordinal classification. In the same manner, the ordinal data problem on kNN and DT has not evolved significantly. Knowing that a DT consistent with the ordinal setting is often desirable to aid decision making, we proposed a strategy based on constraints defined globally over the feature space. This approach was further extended through a bootstrap technique to improve the accuracy of the baseline solution.
Thirdly, we explored a particular problem where in many scenarios there is the opportunity to label critical items for manual revision, instead of trying to automatically classify every item. Therefore, the development of classifiers with an extra output class, the reject class, in-between the decision classes, is attractive where the ordinal problem can easily fit in. We present three new approaches on SOM and a new paradigm initially proposed for the classification of ordinal data to address the classification problem with reject option was delved.
Finally, the proposed methodologies were assessed in two medical applications. "
"Breast Cancer Conservative Treatment (BCCT) has been increasingly used over the last few years as a consequence of its more acceptable cosmetic outcome when compared with mastectomy, but with identical oncological results. Despite significant efforts into the subjective assessment of the aesthetical outcome, breast conservative methods have been used for almost three decades without a standard objective method for aesthetic analysis. Consequently, the application of these treatments produces highly diverse cosmetic results. In order to diminish these dissimilarities, an objective assessment evaluation method is required allowing the correction of working practices and effective comparison of the outcome between centres. Objective evaluation methods have emerged as a way to overcome the poor reproducibility of subjective assessment and have, until few years ago, consisted only of measurements between identifiable points on patient photographs, without any attempt to objectively quantify the overall result. An accessibility increase to new technologies allowed an exploration of new methodologies which led to the creation and development of a new software titled BCCT.core. The automatic localization of all fiducial points would allow the automatic measurement (user-independent) of all the well-known indices correlated with the overall aesthetical result of the BCCT, making cosmetic assessment fast, easy, reliable, and reproducible. This dissertation addresses this need by investigating and proposing several different methods for the automatic detection of fiducial points (nipple complex and breast contour) on the digital photographs of patients. Three new methodologies to detect breast contours with known endpoints were investigated, all incorporating the previous knowledge of breast shapes to improve the detection performance. In the unimodal method it was used the intuitive knowledge that breast contours descend to almost the middle of the breast and then it ascends to the middle of the chest. The prior mask method uses a database of photographs with known breast positions to compute a probability map of the breast contour. Finally, the template method also uses a breast contour database to find a contour appropriate to bias the search that follows. To show the robustness of these methods, they were mutually compared. This resulted in three novel methods for breast contour detection. The automatic detection of the endpoints of the breast contour was also investigated. Armpits were identified as the external endpoints of breast contours whereas the middle of the chest was labelled as the middle point. Several methods have been studied to detect the armpits, with promising results. The novel methodology based on stable paths has shown a noteworthy performance. Nipples are another important point of reference, since their position highly influences the final aesthetic result. In this work a new method is studied for the automatic detection. Finally, several methods were studied in machine learning to obtain the overall aesthetical result. This work was conducted with the simultaneous selection of the most promising features. The methods for the automatic breast contour detection were tested on a set of 190 images, where 120 images had a controlled background and the remaining 70 images had not. When the endpoints were manually positioned these methods attained a mean error of less than 2 centimetres. However, when the endpoints were automatically positioned, the mean error was under 3.5 centimetres. This increase was due to the erroneous detection of the armpit and hence, a new and robust method titled stable paths was introduced. The evaluation of the nipple detection was also performed on this set of images, obtaining a mean error of 5.3 and 4.5 centimetres for the left and right breasts, respectively. The assessment of the overall aesthetical result was based on the Harris four level score. Due to the nature of this problem, two approaches were considered for the evaluation: using classical predictive methods with and without ordinal algorithms. The first obtained a performance of 78.9% whereas the latter obtained a performance of 83%."