Weighting in fuzzy subsumption discovery

We address the problem of discovering fuzzy subsumption relations between concepts from different concept hierarchies. We investigate two approximate reasoning schemes which aim at improving of the fuzzy subsumption relation assessment. The first scheme utilizes the structure of a concept hierarchy and the second takes advantage of a Google-based dissimilarity measure. We present the results of experiments with several music concept hierarchies from actual sites on the Internet. c © Koninklijke Philips Electronics N.V. 2006 iii TN-PR-TN 2006/00495 Philips Restricted Conclusions: In this study, we have addressed the problem of discovering fuzzy subsumption relations between concepts from different concept hierarchies. We have examined two weighting schemes. The first exploits the structure of the CH and the second the NGD dissimilarity measure. From the evidence provided in section 1.4 we can conclude that weighting improves the accuracy of the fuzzy subsumption relation assessment. The structure-based weighting as it is, depends solely on the structure of the CH. Hence, if the CH’s structure does not preserve the actual supergenre-subgenre relations then the weight values can be misleading. Unfortunately, in reality this is the case. ADN, for instance, evenly disperse the classified music genres into a tree-like structure of depth 2. In order to achieve this artificial structure the designers of ADN music metaschema had to sacrifice some supergenre-subgenre relations; Metal and rock are siblings in ADN classification even though metal is a subgenre of rock. Google-based weighting, on the other hand, utilizes the NGD which takes advantage of the vast knowledge available on the web today. Consequently, it should provide a more accurate assessment of the weight values. Therefore, we are inclined to believe that the Google-based weighting is a better method. iv c © Koninklijke Philips Electronics N.V. 2006 Philips Restricted TN-PR-TN 2006/00495