| Cognition intelligence is to make machine have the ability of understanding and thinking like human.For instance,for question “Is Ang Lee a famous director?”,we can give an answer “yes” easily,because we know two prior knowledge that Ang Lee won academy award for best directing,and the award is enough to determine he is a famous director.For the machine,it can get the first prior knowledge from knowledge graph,i.e.,the triple <Ang Lee,award,academy award for best directing>,but there is no way to acquire the second prior knowledge.Concept cognition for knowledge graph is to look at a concept from a comprehensive perspective and clearly grasp its intension,i.e.,the typical characteristics of the things referred by the concept,e.g.,“academy award for best directing” is one of typical characteristics for concept “Famous Director”.Namely,the second prior knowledge can be obtained by concept cognition for knowledge graph that a promising research direction.Different from the existing concept learning and formal concept analysis,which learns and obtains a new concept from data,concept cognition for knowledge graph aim at understanding an existing concept.How to obtain the typical characteristics of the things referred by the concept from the concept related knowledge(such as entity,attribute,attribute value,etc.)in the knowledge graph is the key problem of the concept cognition for knowledge graph.Multi-granularity cognitive computing examines the knowledge implied in data from multiple granular layers,which not only studies cognitive problem by intelligent computing model,but also inspires the design of intelligent computing model by the cognitive law of human brain,so that it is an effective model to realize the concept cognition for knowledge graph.Therefore,based on multigranularity cognitive computing,this thesis realizes the concept cognition for knowledge graph from the perspective of attribute significance and multi-granularity decision rule,which can provide prior knowledge for machine understanding and thinking that is a prerequisite for cognitive intelligence.The main research contents and innovations of this thesis include:First,based on the hierarchical quotient space,a model for determining the siginificance of the concept-relevant attributes is proposed,which realizes the concept cognition for knowledge graph from the perspective of the attribute significance.To solve the problem that different attributes of a concept have different significances for understanding a concept,this thesis realizes the concept cognition for knowledge graph from the perspective of the attribute significance based on hierarchical quotient space.We analyze the difference between general concept cognition and concept cognition for knowledge graph.And then,after fully considering the characteristics of concept-related attributes in knowledge graph,we construct a hierarchical quotient space for the concept in knowledge graph.Finally,we propose using the distance between two hierarchical quotient spaces with and without a certain attribute to determine the siginificance of the concept-relevant attributes.Determining attribute significance can facilitate differentiated treatment of attributes,which can improve understanding concept.Second,based on the multi-granularity knowledge space,a multi-granularity knowledge space model with variable number of layers is proposed to solve the problem that the number of layers of multi-granularity knowledge space and the demand for problem-solving are not synchronized.To address the problem of redundant granular layers and time-consuming transformation between granular layers in the multi-granularity knowledge space,this thesis proposes a multi-granularity knowledge space construction method with variable number of granular layers.We propose a definition of knowledge space distance for the multi-granularity knowledge space with super-subset relationship,and theoretically demonstrate that it is more reasonable than the existing definition of knowledge space distance.And then,a deterministic algorithm and a heuristic algorithm are proposed to select an approximately equidistant k-layer multi-granularity knowledge space.The deterministic algorithm can obtain the optimal result,but its running time is difficult to accept.In contrast,the heuristic algorithm does not obtain the optimal result every time,but we can obtain the optimal result by executing it multiple times,which takes only a little time.The variable k-layer multi-granularity knowledge space can better adapt to the diverse needs of problem solving.Third,the framework of mining the multi-granularity decision rule from the multigranularity information system group is proposed,which mining decision rules from coarser granularity to finer granularity.On the basis of multi-granularity characteristic of attributes and attribute values related to the concept in knowledge graph,this thesis realizes the concept cognition for knowledge graph from the perspective of the multi-granularity decision rule.Conceptrelated attributes/attribute values have four granularities,and information systems according to four granularities can form a multi-granularity information system group,where the coarse-grained results can guide and help mining fine-grained decision rule.In addition,we propose an algorithm of mining multi-granularity decision rule,which constructs a multi-granularity information system group with positive data,then mines frequent maximal attribute pattern from the multi-granularity information system group and derives all frequent attribute patterns.Furthermore,the credible decision rules derived from frequent attribute patterns are verified with negative data.Multi-granularity decision rule provides machine with a more complete understanding of concept from multiple perspectives.Four,based on the three-way decision and the multi-granularity,a multi-granularity three-way decision model is proposed to realize the concept cognition for knowledge graph from the perspective of the multi-granularity decision rule.The method of mining frequent decision rule may miss decision rule with low frequency but high value,so this thesis proposes a new method of mining multigranularity decision rule,which is based on multi-granualrity three-way decision model.The multi-granularity space in existing multi-granularity three-way decision model corresponds to sequential attribute set sequence or attribute hierarchy tree,while the multi-granularity space in this thesis corresponds to different granularities of attribute/attribute value.The multi-granularity three-way decision model divides the granules from coarse granularity into positive granule space,negative granule space and boundary granule space,where only the last one will participate in the three-classification process at finer granularity.After finishing the three-classification processes in four granularities,the multi-granularity positive/negative decision rules are obtained from the positive/negative granule spaces from multiple granularities.Mining multi-granularity decision rule based on multi-granularity three-way decision can avoid missing multigranularity decision rules with low frequency but high value. |
