Слабо-контролируемая групповая классификация

Abstract

В работе решается задача слабо-контролируемого обучения в постановке групповой бинарной классификации. Предполагается, что каждый объект выборки может включать набор подобъектов, относящихся к разным классам. Предлагаемый метод решения основан на выборе информативного признакового пространства и фильтрации обучающей выборки. Описывается применение разработанного метода для прогнозирования степени поражения участков головного мозга у пациентов с ишемическим инсультом по снимкам компьютерной томографии. Weakly supervised learning implies possible uncertainty or fuzziness of the labelling. Current study addresses this problem using the formulation of group binary classification. It is assumed that each sample object may include a set of sub-objects belonging to one of two classes. Objects are described by a set of features; the predicted feature determines the degree to which an object belongs to the “positive” class. It is required to construct a decision function from the training sample to predict the target feature for new objects. The proposed method is based on the selection of informative feature space and filtering the training sample. Both the selection of informative features and the removal of noise observations are carried out on the basis of analysis of the local environment of objects. The degree of similarity between the object and the class is determined by the k nearest neighbours of the object, taking into account their degree of belonging to the target class. For an experimental study of the developed method, the real problem of analyzing tomography images of the brain to predict the degree of damage to its areas in stroke is solved. The results are compared with a number of known methods. A method for constructing a decision function for predicting the degree of belonging of an object to the target class has been developed. The results of an experimental study and comparison with a number of well-known machine learning algorithms (random forest, support vector machine, kNN) confirmed the efficiency of the method for solving the problem of predicting the degree of damage to brain areas in stroke patients. Unlike other similar algorithms, the proposed method allows establishing a set of the most informative features in order to improve the interpretability of the solution and reduce the effect of overfitting.