Semantic Web Recommender Systems 83 Advogato metric [10]. However, the metric can only make boolean decisions with respect to trustworthiness, classifying agents into trusted and untrusted ones. Appleseed [23], our own novel proposal for local group trust computation, allows more fine-grained analysis, assigning continuous trust weights for peers within trust computation range. Rankings thus become feasible. Appleseed's principal concepts de rive from spreading activation models [18], which have been conceived for modelling human semantic memory. Appleseed operates on partial trust graph information, explor ing the social network within predefined ranges only and allowing the neighborhood detection process ro retain scalability. Hereby, high ranks are accorded to trustworthy peers, i.e., those agents which are largely trusted by others with high trustworthiness, similar to PageRank [17]. These ranks are used later on for selecting agents deemed suitable for making recommendations 3.3 Taxonomy.Driven Similarity Metrics Trust allows selecting peers with overall above-average interest similarity. However, for each active user, some peers having completely opposed interests generally exist. The proposition that interpersonal attraction, and hence trust, implies attitudinal similarit does not always hold true. Supplementary filtering thus becomes indispensable. Two approaches are conceivable User-User Closeness. Additional filtering based upon profile similarity c(ai, ai) between agents ai, ai is applied to the neighborhood of ais trustworthy peers. The resulting set of peers only contains trusted peers which are guaranteed to resemble the active user ai Product-USer Relevance. Instead of pruning the very neighborhood, one could also consider the set of all products bk appreciated by at least one of ais neighbors and dispose of those products not fitting the interest profile of ai. We hereby denote the relevance of bk for ai by c(ai, bk) Or For both cases, the filtering task faces the problem of low profile overlap by virtue 3 information sparseness and potentially large product sets [12]. In order to alleviate prevailing issue, we propose taxonomy-driven profile computation [22, 24], which allows to derive similarity between users a; and a, even though these peers have not rated one product in common. Moreover, our novel filtering method also permits to compute similarity between two products bk, bh Profile Generation. Our approach to taxonomy-driven generation of interest profiles [22, 24] extends basic ideas derived from Middleton's ontology-enhanced content-based filtering method [13]. In contrast to generic feature-based filtering, product categories still play an important role, but we have them arranged in a taxonomy and not separate from each other. Products bk bear topic descriptors dke e f(bk) that relate these bk to taxonomic nodes. Several classifications per product are possible, hence f(bk)>1 Each product the user likes infers some interest score for those dke E f(bk). Since these categories dke are arranged in taxonomy C, we can also infer a fractional interest for all super-topics of dke. Hereby, remote super-topics are accorded less interest score than super-topics close to dke. Assume that(p0, P1, .. Pa) gives the path from top elementSemantic Web Recommender Systems 83 Advogato metric [10]. However, the metric can only make boolean decisions with respect to trustworthiness, classifying agents into trusted and untrusted ones. Appleseed [23], our own novel proposal for local group trust computation, allows more fine-grained analysis, assigning continuous trust weights for peers within trust computation range. Rankings thus become feasible. Appleseed’s principal concepts de￾rive from spreading activation models [18], which have been conceived for modelling human semantic memory. Appleseed operates on partial trust graph information, explor￾ing the social network within predefined ranges only and allowing the neighborhood detection process ro retain scalability. Hereby, high ranks are accorded to trustworthy peers, i.e., those agents which are largely trusted by others with high trustworthiness, similar to PageRank [17]. These ranks are used later on for selecting agents deemed suitable for making recommendations. 3.3 Taxonomy-Driven Similarity Metrics Trust allows selecting peers with overall above-average interest similarity. However, for each active user, some peers having completely opposed interests generally exist. The proposition that interpersonal attraction, and hence trust, implies attitudinal similarity does not always hold true. Supplementary filtering thus becomes indispensable. Two approaches are conceivable: – User-User Closeness. Additional filtering based upon profile similarity c(ai, aj ) between agents ai, aj is applied to the neighborhood of ai’s trustworthy peers. The resulting set of peers only contains trusted peers which are guaranteed to resemble the active user ai. – Product-User Relevance. Instead of pruning the very neighborhood, one could also consider the set of all products bk appreciated by at least one of ai’s neighbors and dispose of those products not fitting the interest profile of ai. We hereby denote the relevance of bk for ai by cb(ai, bk). For both cases, the filtering task faces the problem of low profile overlap by virtue of information sparseness and potentially large product sets [12]. In order to alleviate the prevailing issue, we propose taxonomy-driven profile computation [22, 24], which allows to derive similarity between users ai and aj even though these peers have not rated one product in common. Moreover, our novel filtering method also permits to compute similarity between two products bk, bh. Profile Generation. Our approach to taxonomy-driven generation of interest profiles [22, 24] extends basic ideas derived from Middleton’s ontology-enhanced content-based filtering method [13]. In contrast to generic feature-based filtering, product categories still play an important role, but we have them arranged in a taxonomy and not separate from each other. Products bk bear topic descriptors dke ∈ f(bk) that relate these bk to taxonomic nodes. Several classifications per product are possible, hence |f(bk)| ≥ 1. Each product the user likes infers some interest score for those dke ∈ f(bk). Since these categories dke are arranged in taxonomy C, we can also infer a fractional interest for all super-topics of dke . Hereby, remote super-topics are accorded less interest score than super-topics close to dke . Assume that (p0, p1,...,pq) gives the path from top element