Search

Abstract: Numerous industrial and research databases include missing values. It is not uncommon to encounter databases that have up to a half of the entries missing, making it very difficult to mine them using data analysis methods that can work only with complete data. A common way of dealing with this problem is to impute (fill-in) the missing values. This paper evaluates how the choice of different imputation methods affects the performance of classifiers that are subsequently used with the imputed data. The experiments here focus on discrete data. This paper studies the effect of missing data imputation using five single imputation methods (a mean method, a Hot deck method, a Naı¨ve-Bayes method, and the latter two methods with a recently proposed imputation framework) and one multiple imputation method (a polytomous regression based method) on classification accuracy for six popular classifiers (RIPPER, C4.5, K-nearest-neighbor, support vector machine with polynomial and RBF kernels, and Naı¨ve-Bayes) on 15 datasets. This experimental study shows that imputation with the tested methods on average improves classification accuracy when compared to classification without imputation. Although the results show that there is no universally best imputation method, Naı¨ve-Bayes imputation is shown to give the best results for the RIPPER classifier for datasets with high amount (i.e., 40% and 50%) of missing data, polytomous regression imputation is shown to be the best for support vector machine classifier with polynomial kernel, and the application of the imputation framework is shown to be superior for the support vector machine with RBF kernel and K-nearest-neighbor. The analysis of the quality of the imputation with respect to varying amounts of missing data (i.e., between 5% and 50%) shows that all imputation methods, except for the mean imputation, improve classification error for data with more than 10% of missing data. Finally, some classifiers such as C4.5 and Naı¨ve-Bayes were found to be missing data resistant, i.e., they can produce accurate classification in the presence of missing data, while other classifiers such as K-nearest-neighbor, SVMs and RIPPER benefit from the imputation. [Copyright &y& Elsevier]

In this paper, we investigate ways to learn efficiently from uncertain data using belief functions. In order to extract more knowledge from imperfect and insufficient information and to improve classification accuracy, we propose a supervised learning method composed of a feature selection procedure and a two-step classification strategy. Using training information, the proposed feature selection procedure automatically determines the most informative feature subset by minimizing an objective function. The proposed two-step classification strategy further improves the decision-making accuracy by using complementary information obtained during the classification process. The performance of the proposed method was evaluated on various synthetic and real datasets. A comparison with other classification methods is also presented. [ABSTRACT FROM AUTHOR]

Abstract: In multilabel classification tasks the aim is to find hypotheses able to predict, for each instance, a set of classes or labels rather than a single one. Some state-of-the-art multilabel learners use a thresholding strategy, which consists in computing a score for each label and then predicting the set of labels whose score is higher than a given threshold. When this score is the estimated posterior probability, the selected threshold is typically 0.5. In this paper we introduce a family of thresholding strategies which take into account the posterior probability of all possible labels to determine a different threshold for each instance. Thus, we exploit some kind of interdependence among labels to compute this threshold, which is optimal regarding a given expected loss function. We found experimentally that these strategies outperform other thresholding options for multilabel classification. They provide an efficient method to implement a learner which considers the interdependence among labels in the sense that the overall performance of the prediction of a set of labels prevails over that of each single label. [Copyright &y& Elsevier]