Changes 0 . 1 18 . 5 . 04 meh initial setup 1 . 0 15 . 7 . 04 meh completion

ion Matching Figure 3.1: Expertise Based Matching In this section we first introduce a model to semantically describe the expertise of peers and how peers promote their expertise as advertisement messages in the network. Second, we describe how the received advertisements allows a peer to select other peers for a given query based on a semantic matching of query subjects against expertise descriptions. The third part describes how a semantic topology can be formed by advertising expertise. CHAPTER 3. PEER SELECTION 12 3.2.1 Semantic Description of Expertise Peers The Peer-to-Peer network consists of a set of peers P . Every peer p ∈ P has a knowledge base that contains the knowledge that it wants to share. Common Ontology The peers share an ontology O, which provides a common conceptualization of their domain. The ontology is used for describing the expertise of peers and the subject of queries. Expertise An expertise description e ∈ E is a abstract, semantic description of the knowledge base of a peer based on the common ontology O. This expertise can either be extracted from the knowledge base automatically or specified in some other manner. Advertisements Advertisements A ⊆ P × E are used to promote descriptions of the expertise of peers in the network. An advertisement a ∈ A associates a peer p with a an expertise e. Peers decide autonomously, without central control, whom to promote advertisements to and which advertisements to accept. This decision can be based on the semantic similarity between expertise descriptions. 3.2.2 Matching and Peer Selection Queries Queries q ∈ Q are posed by a user and are evaluated against the knowledge bases of the peers. First a peer evaluates the query against its local knowledge base and then decides which peers the query should be forwarded to. Query results are returned to the peer that originally initiated the query. Subjects A subject s ∈ S is an abstraction of a given query q expressed in terms of the common ontology. The subject can be seen a complement to an expertise description, as it specifies the required expertise to answer the query. Similarity Function The similarity function SF : S × E 7→ [0, 1] yields the semantic similarity between a subject s ∈ S and an expertise description e ∈ E. An increasing value indicates increasing similarity. If the value is 0, s and e are not similar at all, if the value is 1, they match exactly. SF is used for determining to which peers a query should be forwarded. Analogously, a same kind of similarity function E × E 7→ [0, 1] can be defined to determine the similarity between the expertise of two peers. Peer Selection Algorithm The peer selection algorithm returns a ranked set of peers. The rank value is equal to the similarity value provided by the similarity function. CHAPTER 3. PEER SELECTION 13 Algorithm 1 Peer Selection let A be the advertisements that are available on the peer let γ be the minimal similarity between the expertise of a peer and the topics of the query. subject := ExtractSubject(query) rankedPeers := ∅ for all ad ∈ A do peer := Peer(ad) rank := SF (Expertise(ad), subject) if rank > γ then rankedPeers := (peer, rank) ∪ rankedPeers end if end for return rankedPeers From this set of ranked peers one can, for example, select the best n peers, or all peers whose rank value is above a certain threshold, etc. 3.2.3 Semantic Topology The knowledge of the peers about the expertise of other peers is the basis for a semantic topology. Here it is important to state that this semantic topology is independent of the underlying network topology. At this point, we don’t make any assumptions about the properties of the topology on the network layer. The semantic topology can be described by the following relation: Knows ⊆ P ×P , where Knows(p1, p2) means that p1 knows about the expertise of p2. The relation Knows is established by the selection of which peers a peer sends its advertisements to. Furthermore, peers can decide to accept an advertisement, e.g. to include it in their registries, or to discard the advertisement. The semantic topology in combination with the expertise based peer selection is the basis for intelligent query routing. 3.3 The Bibliographic Scenario In this section we instantiate the general model for expertise based peer selection from previous section. We use a real-life scenario for knowledge sharing in a Peer-to-Peer environment. In the daily life of a computer scientist, one regularly has to search for publications or their correct bibliographic metadata. Currently, people do these searches with search CHAPTER 3. PEER SELECTION 14 engines like Google and CiteSeer, via university libraries or by simply asking other people that are likely to know how to obtain the desired information. The scenario that we envision here is that researchers in a community share bibliographic metadata via a Peer-to-Peer system. The data may have been obtained from BibTeX files or from a bibliography server such as the DBLP database1. A similar scenario is described in [ANS02], where data providers, i.e. research institutes, form a Peer-to-Peer network which supports distributed search over all the connected metadata repositories. We now describe the bibliographic scenario using the general model presented in the previous section. Peers A researcher is represented by a peer p ∈ P . Each peer has an RDF knowledge base, which consists of a set of bibliographic metadata items that are classified according to the ACM topic hierarchy2. The following example shows a fragment of a sample bibliographic item based on the Semantic Web Research Community Ontology (SWRC)3: The Capabilities of Relational Database Management Systems. Common Ontology The ontology O that is shared by all the peers is the ACM topic hierarchy. The topic hierarchy contains a set, T , of 1287 topics in the computer science domain and relations (T × T ) between them: SubTopic and seeAlso. Expertise The ACM topic hierarchy is the basis for our expertise model. Expertise E is defined as E ⊆ 2 , where each e ∈ E denotes a set of ACM topics, for which a peer provides classified instances. Advertisements Advertisements associate peers with their expertise: A ⊆ P × E. A single advertisement therefore consists of a set of ACM topics for which the peer is an expert on. 1http://dblp.uni-trier.de/ 2http://www.cs.vu.nl/∼heiner/public/SW@VU/classification.daml 3http://ontobroker.semanticweb.org/ontos/swrc.html CHAPTER 3. PEER SELECTION 15 Queries We use the RDF query language SeRQL [BK04] to express queries against the RDF knowledge base of a peer. The following sample query asks for publications with their title about the ACM topic Information Systems / Database Management: CONSTRUCT {pub} {title} FROM {Subject} {}; {title}; {} USING NAMESPACE swrc=, rdf =, acm =, topic= Subjects Analogously to the expertise, a subject s ∈ S is an abstraction of a query q. In our scenario, each s is a set of ACM topics, thus s ⊆ T . For example, the extracted subject of the query above would be Information Systems/Database Management. Similarity Function In this scenario, the similarity function SF is based on the idea that topics which are close according to their positions in the topic hierarchy are more similar than topics that have a larger distance. For example, an expert on ACM topic Information Systems/Information Storage and Retrieval has a higher chance of giving a correct answer on a query about Information Systems/Database Management than an expert on a less similar topic like Hardware/Memory Structures. To be able to define the similarity of a peer’s expertise and a query subject, which are both represented as a set of topics, we first define the similarity for individual topics. [LBM03] have compared different similarity measures and have shown that for measuring the similarity between concepts in a hierarchically structured semantic network, like the ACM topic hierarchy, the following similarity measure yields the best results: