Decision Trees for Uncertain Data

Reviewer: Aris Gkoulalas-Divanis

Value uncertainty is widespread in collected data that contains inaccuracies, such as multiple repeated values or measurement errors. In these cases, typical data mining tasks such as classification may lead to producing erroneous models from the data. Though some have proposed simple methods for dealing with noisy data-for example, replacing the measured value with the average or the median-such methods are too drastic and can greatly influence the accuracy of classification. In this paper, the authors extend classical decision tree building algorithms (based on the framework of C4.5) to cope with uncertain numerical data, where the values of the data items are not crisp, but are described by a range of values leading to a probability density function (pdf). Through extensive experimentation, the authors demonstrate that, when we use suitable pdfs to describe the data items, the accuracy of the decision tree classifiers can be much higher compared to that of classifiers built by using value averages. On the downside, the proposed approaches are much more computationally demanding than traditional approaches, since they require processing of the pdfs. To cope with this limitation, the authors propose three pruning algorithms that can improve efficiency to the point that the execution times become an order of magnitude higher to those of classical decision tree building algorithms. Overall, this interesting paper makes a tangible contribution to the state of the art. Online Computing Reviews Service