Efficient Service Matchmaking using Tree-Structured Clustering Bernhard Tausch 1 , Claudia d’Amato 2 , Steffen Staab 1 , and Nicola Fanizzi 2 1 ISWeb, University of Koblenz-Landau, Germany {tausch,staab}@uni-koblenz.de 2 Department of Computer Science, University of Bari, Italy {claudia.damato, fanizzi}@di.uniba.it Abstract. A tree structure for efficient service matchmaking is created by using a clustering algorithm. Tree nodes represent a superset of all service descriptions in the leaves below. During query processing matchmaking can be restricted to the branches of the tree where tree nodes indicate overlapping between user requests and service descriptions. Good clustering of n service descriptions may improve retrieval time from O(n) to O(log n) for concise queries. 1 Motivation As web services advance from hype to an established technology an increasing number of service implementations emerges. Especially in the context of semantic web services based on description logics (DL) this leads to scalability issues for service discovery. A trade-off has to be found between expressiveness of service descriptions and costs incurred by service matchmaking. This holds in particular for current approaches that perform service matchmaking by DL reasoning as they individually compute the con- sistency of a service request with each service description. To tackle this problem we present an approach that, by clustering available ser- vice descriptions, is able to significantly reduce the number of necessary comparisons for satisfying a request. It is based on abstract service descriptions that represent a set of concrete service executions [2]. By applying an agglomerative algorithm a tree- structured clustering of the service descriptions is obtained. Therein, at each level an intentional description of the nodes (representing description clusters) is constructed. The search space is reduced by matching user requests to node descriptions rather than to all service descriptions, until the leaves of the tree representing the individual service descriptions are reached. During the generation phase of the cluster non-standard DL inference methods are applied before standard DL inferencing is used for matchmaking in the discovery phase. 2 Tree Generation Service descriptions are expressed as ALC concepts. These are clustered into a tree structure where the actual descriptions are represented in the leaves. This clustering is