Search

Classification has been one of the focal points in data mining research and applications. With an effective approach to building compact and accurate associative classification (namely GARC – Gain-based Association Rule Classification (Chen, Liu, Yu, Wei, & Zhang, 2006)) in forms of association rules, this chapter explores a way of fuzzy extension to GARC in dealing with the problem caused by crisp partitions for continuous attribute domains in data. Concretely, the sharp boundaries of the partitioned intervals are smoothened using fuzzy sets (or often conveniently labeled in linguistic terms) so as to reflect a variety of fuzziness on the domains (parameterized in f2), giving rise to a fuzzy associative classifier (i.e., GARCf2). Furthermore, due to the fuzziness involved, the notions of information gain, rule redundancy and conflicts are extended, aimed at providing the desirable features of GARC in the fuzzy extension context for accuracy and compactness. Moreover, data experiments on benchmarking datasets as well as a real-world application illustrate the effectiveness of the proposed fuzzy associative classifier. [ABSTRACT FROM AUTHOR]

Given a set of graphs with class labels, discriminative subgraphs appearing therein are useful to construct a classification model. A graph mining technique called Chunkingless Graph-Based Induction (Cl-GBI) can find such discriminative subgraphs from graph structured data. But, it sometimes happens that Cl-GBI cannot extract subgraphs that are good enough to characterize the given data due to its time and space complexities. Thus, to improve its efficiency, we propose pruning methods based on the upper-bound of information gain that is used as a criterion for discriminability of subgraphs in Cl-GBI. The upper-bound of information gain of a subgraph is the maximal one that its super graph can achieve. By comparing the upper-bound of each subgraph with the best information gain at the moment, Cl-GBI can exclude unfruitful subgraphs from its search space. Furthermore, we experimentally evaluate the effectiveness of the pruning methods on a real world and artificial datasets. [ABSTRACT FROM AUTHOR]