Multiresolution Knowledge Mining using Wavelet Transform
R. Pradeep Kumar, Panduranga Naidu Nagabhushan · Engineering letters · 2007
Most research in Knowledge Mining deal with the basic models like clustering, classification, regression, association rule mining and so on. In the process of quest for knowledge most of the knowledge mining algorithms end up in generating global knowledge while losing focus on the local knowledge. This happens oftenly due to two reasons. First reason is due to dimensionality reduction. The problem of dimensionality reduction has been viewed as the reduction of features to the maximum extent possible while being able to retain the information conveyed by the data set. But most of the dimensionality reduction techniques reduce the dimensions keeping only the retention of global knowledge in mind while compromising with the loss of local knowledge. Second reason is due to optimized feature selection for making a global classification while not being bothered about the intra class relationship. In this paper we present methodologies using wavelet transform for overcoming the loss of local knowledge along the process of mining. First we propose a discrete wavelet transform based multi resolution approach to capture the local knowledge along the process of dimensionality reduction and also being capable of representing the global knowledge with few number of wavelet coefficients. A comparison with PCA has also been made here, which strongly supports our technique using discrete wavelet transform to produce more number of local knowledge and also no/minimum misleading information. Secondly we propose continuous wavelet transform based multiresolution approach for knowledge mining through a novel histogram distance measure. Along both this approach we also propose methodologies for capturing knowledge packets from different sources and integrating this knowledge for generating a comprehensive knowledge base. In the first part knowledge packets are generated along individual sources, filtered and then integrated. In the second part, summing the distances between the objects when observed from different sources does the integration and then the knowledge is mined from overall distance matrix to obtain comprehensive knowledgebase. These foundational techniques are illustrated with a set of '8-O-X' spatial dataset.