Research On Domain Knowledge Based Knowledge Discovery

Currently, Research and application of Knowledge Discovery is a hot topic among scientific research and real application. But when facing the always domain-related complicated application environments, Current research fail to achieve satisfying results because they can not incorporate domain knowledge effectively. How to realize domain knowledge based knowledge discovery from enduser's prespective is a hot focus among scientific research and real application. With the rapid development of Service Oriented Architecture and Semantic Web, how to perform service oriented knowledge discovery is a key topic of next generation of knowledge discovery systems. So this thesis is focused on domain knowledge based knowledge discovery, more concretely, research work was carried out around service oriented knowledge discovery architecture and domain knowledge based knowledge discovery algorithms. Following innovative contributions were achieved:1. This thesis proposed a novel service oriented architecture for knowledge discovery-SOA4KD, which embrace a quality evaluation architecture for knowledge discovery service, and an IBL(Instance Based Learning) based meta-learning algorithm is proposed to ensure the quality of knowledge discovery services.2. Based on ontology services, this thesis proposed an approach to enable end-user to input their knowledge discovery requirements using natural language. To achieve this goal, the domain ontology is categorized in further detail, and then an extended knowledge discovery task ontology is proposed, which link the possible problem elements and knowledge discovery tasks. Finally, a semantic interpretation algorithm is proposed to transfer the user requirements expressed with natural language to computer-apprehensible formal language, and the algorithm is proved to be sound and complete. This approach enables end-user to self-build knowledge discovery application on SOA4KD.3. This thesis proposed a novel kNN approach based on semanticdistance------SDkNN. This approach analysis the semantic difference between valuesof an attribute and presents how to calculate the semantic distance based on domain ontologies services, the semantic distance is then used to improve the traditional kNN methods. Experiments on UCI(University of California, Irvine) machine learning repository and real application datasets show that the overall performance of SDkNN outperforms the traditional one, especially when the data is incomplete. SDkNN also has the desirable application value in practice.4. This thesis proposed a method for discovering novel multi-level classification rules using ontology and domain knowledge service. Two heuristic rules are proposed for determining whether an attribute should be divided further and knowledge is pushed layer by layer. An algorithm for discovering unexpected multi-level classification rules using ontology is presented. Experiment shows the feasibility of the method.5. This thesis proposed a domain-knowledge based two-stage neural network model. The contribution is composed of two parts. Firstly, traditional RBFNN model need to determine the numbers and centers of neurons manually and the output of traditional model is always constrained by the intial distribution of source data and is unsatifying especially on unbalanced data. This thesis proposed a VSOM based two-stage neural network model. Experiments show this model can solve above problem effectively. Secondly, based on abovementioned model, an insurance loss prediction model incorprating domain knowledge is proposed, terrain factor based on Digital Elevation Model(DEM) is abstracted as input, an approach is proposed to incorporate local domain knowledge(i.e. flood risk map) with the output results, which improve the generation ability of the model. This is also proved by experiments.6. A prototype system of service oriented knowledge discovery system based on Model-View-Controller design pattern is implemented. The application potential is demonstrated by real application among insurance comp...