Research On Ontology Learning Algorithm Based On Graph Structure

The term ontology comes from the field of philosophy,which means the description of the essence of things.Ontology,a structured conceptual representation model,has been introduced into a wide range of fields both in natural and social science,with the development of computer theory and technology.As a tool for concept representation,storage,query and analysis,ontology is applied widely in the fields of genetics,botany,nutrition and other fields,and providing scholars with convenience in various researches and applications.The deepening understanding of essence of things has been leading to the more diversity of the concepts and corresponding attributes and the more complexity of their corresponding ontology graph structures.However,most of the ontology similarity algorithms are still heuristic,that is,the similarity calculation formula of the ontology concepts is defined artificially based on the characteristics of the ontology.The biggest drawback of this trick is that it relies heavily on the knowledge of domain experts,and the similarity calculation formulas given intuitively often can't exactly describe the intrinsic correlation of the ontology data.Therefore,it is impossible to measure the similarity between ontologies.The paper tends to consider the problems mentioned above and try to obtain the ontology similarity calculation method in view of machine learning.The research of this thesis is based on ontology graph models,and obtain the similarity measuring by means of ontology learning function learning.Furthermore,the theoretical analysis of these ontology learning algorithms are produced.The results of main research are as follows:(1)Under two hypothesis of IID and non-IID ontology data,the analysis of uniform stability and error bounds are given respectively.In IID setting,the error bound and related theoretical analysis under the hypothetical space stability framework are given;while under the condition of non-IID,specific theoretical analysis is given for the decision graph framework,and the corresponding learning statistics characteristics are obtained.Specifically,in non-IID and large samples setting,the replacement and deletion of a single sample is extended to replace and delete multiple samples,and the upper bound estimate of the corresponding error bounds are determined.(2)In view of the limitations of the existing ontology learning algorithm pairwise comparison,by forming a hyperedge from multiple ontology vertices in each group,the hypergraph model is applied to the ontology similarity calculation,and the weight of each vertex is calculated by a random path.The corresponding deduced bipartite graph is obtained,and the hypergraph baded ontology learning algorithm is determined.(3)The multi-dividing ontology learning algorithm is proposed based on the tree structures of the ontology graph,The ontology vertices under the root vertex is devided into several classes according to branches,and gives the order relationship between each class.In the learning process,the values of the ontology vertices from different classes are compared in pairs,and the comparison conclusion is required to conform to the pre-set category order setting.The ontology learning model is obtained by sparse vector calculation,and theoretical analysis shows that the algorithm can be regarded as a support vector machine based multi-dividing ontology learning algorithm.(4)From different perspectives of the samples,four types of multi-dividing ontology learning algorithms are proposed:(i)Aiming at scenarios where the ontology sample capacity and storage space are unlimited,a multi-dividing ontology learning algorithm under the condition of super large samples is proposed.The statistical tricks are used to analyze this kind of algorithm,and its generalized bound is obtained theoretically;(ii)A multi-dividing ontology learning algorithm in small sample setting is proposed.It extracts a sub-sample from the original large sample,and then reduces the calculation complexity of the algorithm by reducing the sample capacity;(iii)Using the two-sample learning trick,a multi-dividing ontology learning algorithm is obtained,each sub-class sample is divided into two categories,and the optimal ontology function is deduced by minimizing the two-sample error;(iv)A multi-dividing ontology learning algorithm is determined based on buffer update,which mainly considers how to save storage space in the process of ontology learning under the condition of limited storage space.(5)From the perspective of ontology function,a multi-segment ontology learning algorithm based on weak ontology function is proposed.In addition,since the number of classes of vertices in tree-shape ontology graphs are very small in the practical applications,it is pointed out that the original multi-dividing learning algorithm can be further improved by modifying the truth function,and a specific modification trick is given.This thesis contains 7 figures,5 tables and 140 references.