Discovery of the Latent Supportive Relevance in Medical Data Mining

The purpose of a classification learning algorithm is to accurately and efficiently map an input instance to an output class label, according to a set of labeled instances. Decision tree is such a method that most widely used and practical for inductive inference in the data mining and machine learning discipline (Han & Kamber, 2000). However, many decision tree learning algorithms degrade their learning performance due to irrelevant, unreliable or uncertain data are introduced; or some focus on univariate only, without taking the interdependent relationship among others into consideration; while some are limited in handling the attributes1 with discrete values. All these cases may be caused by improper pre-processing methods, where feature selection (FS) and continuous feature discretization (CFD) are treated as the dominant issues. Even if a learning algorithm is able to deal with various cases, it is still better to carry out the pre-processing prior the learning algorithm, so as to minimize the information lost and increase the classification accuracy accordingly. FS and CFD have been the active and fruitful fields of research for decades in statistics, pattern recognition, machine learning and data mining (Yu & Liu, 2004). While FS may drastically reduce the computational cost, decrease the complexity and uncertainty (Liu & Motoda, 2000); and CFD may decrease the dimensionality of a specific attribute and thus increase the efficiency and accuracy of the learning algorithm. As we believe that, among an attributes space, each attribute may have certain relevance with another attribute, therefore to take the attributes relevant correlation into consideration in data pre-processing is a vital factor for the ideal pre-processing methods. Nevertheless, many FS and CFD methods focus on univariate only by processing individual attribute independently, not considering the interaction between attributes; this may sometimes loose the significant useful hidden information for final classification. Especially in medical domain, a single symptom seems useless regarding diagnostic, may be potentially important when combined with other symptoms. An attribute that is completely useless by itself can provide a significant performance improvement when taken with others. Two attributes that are useless by themselves can be useful together (Guyon & Elisseeff, 2003; Caruana & Sa, 2003). For instance, when learning the medical data for disease diagnostic, if