The Research Of Fusion Method For Relational Domain Knowledge Oriented To Data Mining

Current data mining methods mostly focus on the raw data, which have no domain knowledge fusion and are dependent on human intervention for completing user participation in the process of knowledge discovery. While in practice, data vary in levels, with some raw, and some closely related to others. Combination of relevant data and granularities can better reveal the inner rules. Therefore, fully exploiting the relevant domain knowledge in the process of data mining can "observe and analyze the same problem from different granularities",hence acquire knowledge at different data levels, and obtain flexible conversion at different levels without difficulty. This is now an important research program.In application, most data extend to domain knowledge based on attributes, and the corresponding domain knowledge is shown through relation tables. Therefore, this paper focus on representation of relational knowledge domain and its fusion with data mining research, so as to automatically detect knowledge.The main works done in the dissertation are summarized as follows:(1) Structured representation model of relational domain knowledge (DKMRM,Domain Knowledge of Multi-Relations Model) is proposed. Depending on the relation mapping or projection of the attributes of domain knowledge, contextual relation tables are established and accordingly complex representation of multi-relational system is established. This model together with necessary transformation strategies can facilitate the user retrieval of personalized knowledge for it can generalize or specify the raw data to a rational level.(2) DKMRM based on relational domain knowledge is applied to data mining, and data mining-oriented fusion method for relational domain knowledge is proposed.Classification problem is employed in application, and corresponding classification data mining framework is constructed. Compared to regular mining methods, this method can effectively solve the key problems in transfer source, transfer path, termination strategy, and transfer deviation statistics.(3) For different data levels, algorithm CC-DKMR(Classification of Characters based on Domain Knowledge of Multi-Relations) is proposed based on attribute selection of multi-relational classification mining, and algorithm CS-DKMR(Classification of Sheets based on Domain Knowledge of Multi-Relations) is proposed based on multi-relational table selection. These algorithms seek mining model and flexible conversion mechanism in different hierarchy of data granularities, from the domain knowledge to obtain more valuable knowledge. Experiments have proved the effectiveness of the method.(4) 'Oracle' problem is common in data mining algorithms (programs), which invalidates the traditional evaluation methods. In the evaluation stage of data mining,the proposed domain knowledge evaluation method is based on metamorphic testing technology, which covers the analysis of structure metamorphic relations of classification rules, association rules and clustering analysis algorithm. The experimental results show that this method can effectively achieve the evaluation purpose, and is applicable for broader use in other fields.