| Abstract | A semantically enhanced weighted tree similarity algorithm for buyer-seller match-making is presented. First, our earlier global (structural) similarity measure over (product) partonomy trees is enriched by taxonomic semantics: Inner nodes can be labeled by classes whose partial subsumption order is represented as a background taxonomy tree that is used for class similarity computation. In particular, the similarity of any two classes can be defined via the weighted length of the shortest path connecting them in that taxonomy. To enable similarity comparisons between specialized versions of the background taxonomy, we encode subtaxonomy trees into partonomy trees in a way that allows the direct reuse of our partonomy similarity algorithm and permits weighted (or ‘fuzzy') taxonomic subsumption with no added effort. Second, leaf nodes can be typed and each type be associated with a local, special-purpose similarity measure realizing the semantics to be invoked when computing the similarity of any two of its instances. We illustrate local similarity measures with e-Business types such as “Currency”, “Address”, “Date”, and “Price”. For example, the similarity measure on “Date”-typed leaf node labels translates various notations for date instances into a normal form from which it linearly maps any two to their similarity value. Finally, previous adjustment functions, which prevent similarity degradation for our arbitrarily wide and deep trees, are enhanced by smoother functions that evenly compensate intermediate similarity values. |
|---|
