Approaches To Knowledge Discovery And Their Applications Via Axiomatic Fuzzy Sets And Knowledge Graph Theory

Axiomatic Fuzzy Sets (AFS) theory provides a new method to research human cogni-tion, natural language semantic and thinking logics starting from the fuzziness of human perceptions and cognition. AFS has been applied to many fields such as knowledge rep-resentation, pattern recognition, clustering analysis and formal concept analysis and so on. Knowledge graph is a new method for knowledge representation which establishes semantic interpretation models based on psychology and philosophy. Knowledge graph has been applied to decision support systems, model expert systems, machine translation as well as to describe texts. This thesis focuses on fuzzy classification association rule mining and the stability of fuzzy rule base which are often encountered in knowledge discovery and representation based on AFS and knowledge graph theory. Furthermore, some problems about the coding and recognition of human body plane motion are also studied in the thesis. Main topics include:1. This thesis proposes an association classification algorithm namely AFSRC real-ized in the framework of AFS theory, which provides a simple and efficient rule generation mechanism. Classical interesting measure namely "Support" prefers the majority class. As a result the rules for minor class would be abandoned. A new concept of "Fuzzy Class Support" is proposed to coped with this problem, which can retain meaningful rules for imbalanced classes. This paper also gives the concept of "optimal fuzzy confidence trun-cation" with its computation method which can get over the difficulty of selecting the minimal confidence as well as to keep the different size rules with higher fuzzy confidence. AFSRC can handle different data types and produce membership functions automatically by processing available data. In addition, AFSRC synthesizes various knowledge discovery techniques such as feature selection, data reduction and post pruning. The experimental results show that AFSRC outperforms most of other methods and forms a classifier with high accuracy and more interpretable rule base of smaller size while retaining a sound balance between these two characteristics. 2. Many concepts have alot of different definitions in natural language. For extract-ing substantive information and words from the various definitions of one concept, this thesis extracts representative words by choosing synonyms and establishes word graph for concepts based on knowledge graph theory. By a careful analysis of immanent informat ion of word graph, some representative features are summarized. Fuzzy rules are extracted from AFS fuzzy decision trees which are established on the information of vertices and edges of word graphs. Then the decision about important vertices and edges can be made. The detail process of the decision-making on the example concept "democracy" is pro-vided and the results demonstrate the validness of the proposed method. The proposed method can be used to acquire the main words of consistency definition of a concept as well as to simplify the word graph.3. By an exhaustive study of the properties of rules and relationships between them, a semantic lattice of fuzzy rule base is established in the framework of AFS theory. All the fuzzy rules obtained by different algorithms or training data sets under the same fuzzy partition can be put in the same semantic lattice which offers a new mathematics tool for analyzing fuzzy rules. The stability of fuzzy rule bases is analyzed from three aspects through semantic lattice. Furthermore, definition of "index" of a fuzzy rule base is introduced to describe the compactness of the semantic lattice formed by the fuzzy rule bases. The proposed semantic lattice method can be used to select rule based classifiers as well as to develop ensemble learning classifier. The experiments verify the validity of the proposed measurements.4. Considering motions such as simple shifts and complex turns, five types of coding methods for human upperbody plane motions are proposed. One of the coding methods is to take O or1as components of a code vector with four components, which can classify the movements into16standard states. Combined with the priori knowledge, these standard codes can be used to recognize actions and to compress image information. Two methods to measure the similarity of two motions based on proposed codings are also given. A measure called agitation, which can be seen as an indicator for a change between motions, is also introduced and it is proved to be valid. The encodings are tested on data in the form of measurements of body motions obtained from vidcoconference. The results show that it can recognize specific movement in the videoconference.