A Graph Analytics Framework for Knowledge Discovery

Abstract

In the current data movement, numerous efforts have been made to convert and normalize a large number of traditionally structured and unstructured data to semi-structured data (e.g., RDF, OWL). With the increasing number of semi-structured data coming into the big data community, data integration and knowledge discovery from heterogeneous do mains become important research problems. In the application level, detection of related concepts among ontologies shows a huge potential to do knowledge discovery with big data. In RDF graph, concepts represent entities and predicates indicate properties that connect different entities. It is more crucial to figure out how different concepts are re lated within a single ontology or across multiple ontologies by analyzing predicates in different knowledge bases. However, the world today is one of information explosion, and it is extremely difficult for researchers to find existing or potential predicates to per form linking among cross domains concepts without any support from schema pattern analysis. Therefore, there is a need for a mechanism to do predicate oriented pattern analysis to partition heterogeneous ontologies into closer small topics and generate query to discover cross domains knowledge from each topic. In this work, we present such a model that conducts predicate oriented pattern analysis based on their close relationship and generates a similarity matrix. Based on this similarity matrix, we apply an innovative unsupervised learning algorithm to partition large data sets into smaller and closer topics that generate meaningful queries to fully discover knowledge over a set of interlinked data sources. In this dissertation, we present a graph analytics framework that aims at providing semantic methods for analysis and pattern discovery from graph data with cross domains. Our contributions can be summarized as follows: • The definition of predicate oriented neighborhood measures to determine the neighborhood relationships among different RDF predicates of linked data across do mains; • The design of the global and local optimization of clustering and retrieval algorithms to maximize the knowledge discovery from large linked data: i) top-down clustering, called the Hierarchical Predicate oriented K-means Clustering;ii)bottom up clustering, called the Predicate oriented Hierarchical Agglomerative Clustering; iii) automatic topic discovery and query generation, context aware topic path finding for a given source and target pair; • The implementation of an interactive tool and endpoints for knowledge discovery and visualization from integrated query design and query processing for cross do mains; • Experimental evaluations conducted to validate proposed methodologies of the frame work using DBpedia, YAGO, and Bio2RDF datasets and comparison of the pro posed methods with existing graph partition methods and topic discovery methods. In this dissertation, we propose a framework called the GraphKDD. The GraphKDD is able to analyze and quantify close relationship among predicates based on Predicate Oriented Neighbor Pattern (PONP). Based on PONP, the GraphKDD conducts a Hierarchical Predicate oriented K-Means clustering (HPKM) algorithm and a Predicate oriented Hierarchical Agglomerative clustering (PHAL) algorithm to partition graphs into semantically related sub-graphs. In addition, in application level, the GraphKDD is capable of generating query dynamically from topic discovery results and testing reachability be tween source target nodes. We validate the proposed GraphKDD framework through comprehensive evaluations using DBPedia, Yago and Bio2RDF datasets.