earticle

영어

Data mining is the process of extracting useful and yet unknown information such as patterns or associations hidden in stored data. Among various existing techniques applied to search for interesting patterns, decision tree is one of the most popular tools used for data mining. Most data mining techniques are data-driven, however, data repositories of interest in data mining applications can be very large and noisy. Noise is a random error in data. Noise in a data set can happen in different forms: misclassification or wrong labeled instances, erroneous or distorted attribute values, contradictory or duplicate instances having different labels. All kinds of noise can more or less affect the learning performance. The most serious effect of noise is that it can confuse the learning algorithms to produce complex and distorted results. The long and complex results are due to the attempt to fit every training data instance, including noisy ones, into the concept descriptions. This is a major cause of overfitting problem. Most learning algorithms are designed with the awareness of overfitting problem due to noisy data. Prepruning and postprocessing are two major techniques applied to avoid growing a decision tree too deep down to cover the noisy training data. These techniques are tightly coupled to the tree induction phase. We, on the contrary, design a loosely coupled approach to deal with noisy data. Our noise-handling feature is in a separate phase from the tree induction. Both corrupted and uncorrupted data are clustered and heuristically selected prior to the application of tree induction engine. We observe from our experimental study that tree models produced from our approach are as accurate as the models generated by conventional decision tree induction approach. Moreover, upon highly corrupted data our approach shows a better performance than the conventional approach.