Study On Approaches To Ontology Learning Oriented Granular Computing

With the development of web, there are an increasing number of subject-oriented domain data sources (the domain information system in this dissertation) with substantial information. This kind of data source is semi-structured, which has characteristics of incomplete internal structure, incomplete, imprecise or uncertain contents, a mass of data, dynamic and distributed storage and so on. Approaches to implementing knowledge extraction and knowledge representation from this kind of data source have become much more important now.However, the existing ontology learning methods are typically proposed based on unstructured and structured data sources, whereas corresponding treatment of ontology learning for semi-structured data sources usually bases on unstructured methods, which fail to fully consider the hidden structural features of data sources. The existing literature seldom covers ontology learning methods and technologies based on semi-structured data sources. Therefore, with this kind of semi-structured data of the domain information system, big challenges await ontology learning algorithms.On the other hand, although granular computing (GrC) methods in data mining have been explored widely and deeply by many scholars, there are few based on data source in the domain information system in data mining. It is important to research approaches to ontology learning oriented granular computing, which has theoretical significance and application prospect in that it will not only extend theories and methods of data mining and knowledge discovery, but also provide an effective way for ontology learning in complicated data sources.This dissertation is set in the domain information system with the guideline of the GrC theory, focusing on GrC data mining and oriented toward ontology learning, aiming to carry on some explorative work in this field.What has been explored in this dissertation are as follows:(1) Through the expansion of information functions in the domain information system, formal description and corresponding rough sets of the system are given, and then the domain concept granular space model oriented ontology learning is proposedFirstly, the domain information system should expand information functions, due to the fact that the system is semi-structured data with the features of incomplete and multi-valued data and so on. Through the expansion of information functions, various relations among objects in the domain information system, including equivalent relation, can be defined, so corresponding rough set model is proposed, and application environment of rough set is extended.Secondly, to meet the need of concept and taxonomic relation learning in ontology learning, one method of hierarchy granulating with respect to the domain information system is proposed. The method can generate information granules of different abstract degrees (referred to as object granules in this dissertation). Feature descriptions of object granules and corresponding feature support sets are defined from the approximate concept of GrC, and the domain object granular space is constructed. Concept granule is defined by combining object granule with its corresponding feature description, which represents the domain (approximate) concept. Then the domain object granular space, which is generated by hierarchy granulating, induces the corresponding domain concept granular space. The relationship and operation between object granules, corresponding properties and the representation of granule approximation in the domain uncertain information are discussed, and some features of concept granular space model are analyzed. And, ontology concept learning and taxonomic relation learning are formulated as the generating process of the domain concept granular space based on the domain information system.(2) In response to the need of acquiring concepts and their taxonomic relation in ontology learning, a multi-hierarchy concept acquisition approach based on granular computing is proposedGiven the domain concept granular space model and the situation of uncertain or imprecise characteristic value in the domain information system, the granulating criterion is defined and the domain multi-hierarchy concept acquisition algorithm (CGS) is proposed. Therefore it provides one feasible method and technology for ontology concept learning and taxonomic relation learning of the domain information system. And based on the algorithm, the increment method of the domain multi-hierarchy concept acquisition (CGS2) is proposed according to the dynamic and distributed characteristics on the domain information system, thus it can be adapted to the needs of dynamic data sources effectively. Through the comparison of algorithm testing, prototype demonstration testing and preliminary application of prototype demonstration system of ontology learning, the proposed methods are proven to be effective.(3) In response to the need of relation concept acquisition in ontology learning, a granular computing method for mining association relation with domain multi-dimension and multi-level is proposedIt is often done to attain relation concept and non-taxonomic relation in ontology learning through the learning of association relation. But this method finds association relation between concepts based on concept sets. This dissertation puts forward an approach (G-Approach) to mining domain multi-dimension and multi-level of association relation based on granular computing by employing characteristics range concept hierarchy and optimization strategy of association relation mining. Some real examples are provided to dwell on the proposed algorithm, choosing different types of data sets and other typical mining methods to test and compare from different aspects, and the test results testify to the feasibility of the approach. Besides, the proposed approach is running on the domain information system directly, and association relation between different concept hierarchies can be found. Therefore, the approach not only extends the data mining method to complicated data sources, but also provides one workable solution and technology of elation concept learning in ontology learning.Meanwhile, an increment approach (G-Approach2) to mining association relation with domain multi-dimensional and multi-level based on G-Approach is proposed, which can cater to the need of dynamic or distributed domain information system. The results of comparison and testing show that capability of G- Approach2 is better than that of G-Approach in terms of non dense data source, while the test results on completed dense data sets show the CGS method is much better. Therefore, we should explore further to improve the increment method targeted at dense data sets.(4) In response to the need of ontology non-taxonomic relation learning, the acquisition method of cross relationship between concept granules based on different concept granular spaces is proposedGiven different aspects or sides constructing different domain concept granular spaces from the same domain information system, an approach to acquiring cross relationship between concept granules is proposed based on the context analysis between concept granules. Therefore, it provides one feasible method and technology of ontology non-taxonomic relation learning based on the domain information system.(5) Ontology learning framework based on granular computing is proposedSimple ontology generating algorithm based on granular computing combined with formal concept lattice is proposed, and a framework for ontology learning based on granular computing is also established. The framework mainly includes three parts- domain concept acquisition, domain concept relation acquisition, and mapping of the concepts and relations to ontology class and relationship, including class mapping, relationship mapping and instance mapping and so on. Then according to the approaches proposed to ontology learning oriented granular computing and the framework, corresponding prototype demonstration system is designed, the proposed framework and approaches are effective proved by operation testing on the given data set.