Robust Data Mining: An Integrated Approach

The continuous improvement and application of information system technologies have become widely recognized by the industry as critical for maintaining a competitive advantage in the marketplace (Shin et al., 2006). It is also recognized that improvement and application activities are the most efficient and cost-effective when implemented during an early process/product design stage. Data mining (DM) has emerged as one of the key features of many applications in computer science. Often used as a means for predicting the future directions and extracting hidden limitations and specifications of a product/process, DM involves the use of data analysis (DA) tools to discover previously unknown and valid patterns and relationships from a large database. Most DM methods for factor selection reported in literature may yield a number of factors associated with interesting response factors without providing detailed information, such as relationships between the input factor and response, statistical inferences, and analyses (Yang et al., 2007; Witten & Frank, 2005). Based on this, Gardner and Bieker (Gardner & Bieker, 2000) suggested an alternative DA approach toward resolving semiconductor manufacturing problems in order to determine the significant factors. Furthermore, Su et al. (Su et al., 2005) developed an integrated procedure combining a DM method and Taguchi methods. DA is a term coined to describe the process of sifting through large databases for discovering interesting patterns and relationships. This field spans several disciplines such as databases, machine learning, intelligent information systems, statistics, and expert systems. Two approaches that enable the application of standard machine learning algorithms to large databases are factor selection and sampling. Factor selection is known to be an effective method for reducing dimensionality, removing irrelevant and redundant data, increasing mining accuracy, and improving result comprehensibility (Yu & Liu, 2003). Consequently, factor selection has been a fertile field for research and development since the 1970s and proven to be efficient in removing irrelevant and redundant features, increasing efficiency in mining tasks, improving mining performance like predictive accuracy, and enhancing comprehensibility of the learned results. The factor selection algorithm performs a search through the space of feature subsets (Allen, 1974). In general, two categories of the algorithm have been proposed to resolve the factor selection problem. The first category is based on a filter approach that is independent of the learning algorithms and serves as a filter to sieve out the irrelevant factors. The second category is based on a wrapper approach, which uses an induction algorithm itself as part of the function evaluating the factor subset (Langley, 1994). Since most of the filter methods are based on a heuristic O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg