MapPSO Results for OAEI 2008

We present first results of an ontology alignment approach that is based on discrete particle swarm optimisation. In this paper we will firstly describe, how the algorithm approaches the ontology matching task as an optimisation problem, and briefly sketch how the specific technique of particle swarm optimisation is applied. Secondly, we will briefly discuss the results gained for the Benchmark data set of the 2008 Ontology Alignment Evaluation Initiative. 1 Presentation of the system We introduce the Ontology Mapping by Particle Swarm Optimisation (MapPSO) system as a novel research prototype, which is expected to become a highly scalable, massively parallel tool for ontology alignment. In the following subsection the basic idea of this approach will be sketched. 1.1 State, purpose, general statement The MapPSO algorithm is being developed for the purpose of aligning large ontologies. Instance mapping however is not part of our efforts. Motivated by the observation that ontologies and schema information such as thesauri or dictionaries are not only getting numerous on the web, but also are becoming increasingly large in terms of the number of classes/concepts and properties/relations. This development raises the need for highly scalable tools to provide interoperability and integration of various heterogeneous sources. On the other hand the emergence of parallel architectures provide the basis for highly parallel and thus scalable algorithms which need to be adapted to these architectures. For the presented MapPSO method we formulated the ontology alignment problem as an optimisation problem which allowed us to employ a discrete variant of particle swarm optimisation [1, 2], a population based optimisation paradigm inspired by social interaction between swarming animals. Particularly the population based structure of this method provides high scalability on parallel systems. Particle swarm optimisation furthermore belongs to the group of anytime algorithms, which allow for interruption at any time and will provide the best answer being available at that time. Particularly this property might be interesting when an alignment problem is subject to certain time constraints. 1.2 Specific techniques used MapPSO utilises a discrete particle swarm optimisation (DPSO) algorithm, based in parts on the DPSO developed by Correa et al. [1, 2], to tackle the ontology matching problem as an optimisation problem. The core element of this optimisation problem is the objective function which supplies a fitness value for each candidate alignment. To find solutions for the optimisation problem, MapPSO simulates a set of particles whereby each particle is a candidate alignment comprising a set of initially random mappings3. Each of these particles maintains a memory of previously found good mappings (personal best) and the swarm maintains a collective memory of the best known alignment so far (global best). In each iteration, particles are updated by changing their sets of correspondences in a guided random manner. Correspondences which are also present in the global best set are more likely to be kept, as are those with a very good evaluation. In addition the number of correspondences represented by each particle also changes according to the number of correspondences in the global best alignment in a self-adaptation process. Each candidate alignment of two ontologies is scored based on a weighted sum of quality measures of the single correspondences, and the number of correspondences it consists of. The currently best alignment is the one with the best known fitness rating according to these criteria. According to this revisit of the ontology matching problem, a particle swarm can be applied to search for the optimal alignment. For each correspondence the quality score is calculated based on an aggregation of scores from a configurable set of base matchers. Each base matcher provides a distance measure for each correspondence. Currently the following well known base matchers are used: – SMOA string distance [3] for entity names – SMOA string distance for entity labels – WordNet distance for entity names – WordNet distance for entity labels – Vector space similarity [4] for entity comments – Hierarchy distance to propagate similarity of superclasses / superproperties – Structural similarity of classes derived from properties that have them as domain or range classes – Structural similarity of properties derived from their domain and range classes For each correspondence the available base distances are aggregated by applying the OWA operator [5]. The OWA operator performs an Ordered Weighted Average aggregation of the base distances by ordering the base distances and applying a fixed weight vector. The evaluation of the overall alignment of each particle is computed by aggregating all its correspondence distances and accounting for the number of correspondence represented by this particle. In the current implementation each of the particles runs in an individual thread and all fitness calculations and particle updates are performed in parallel. The only sequential portion on the algorithm is the synchronisation after each iteration to acquire the fitness value from each particle and determine the currently global best alignment. 3 Currently only 1:1 alignments are supported. 1.3 Adaptations made for the evaluation Since MapPSO is an early prototype, we did use the OAEI 2008 Benchmark test data during the development process. No specific adaptations have been made. 1.4 Link to the system and parameters file The release of MapPSO for OAEI 2008 is located in the package MapPSO at http://ontoware.org/projects/mappso/ 1.5 Link to the set of provided alignments (in align format) The alignment results of MapPSO for the Benchmark test case of OAEI 2008 are located in the package alignResults at http://ontoware.org/projects/mappso/