Finding Patterns in Semantic Graph Formalisms

When intelligence analysts are required to understand a complex uncertain situation, one of the techniques they use most often is to simply draw a diagram of the situation. The diagrams, also called attributed relational graphs or semantic graphs, generally capture the meaning about the situation in their nodes and edges, where the nodes represent concepts/entities and the edges represent the relations/connectivity between the nodes. An important research problem in the area of semantic knowledge discovery and pattern analysis is to identify common/uncommon patterns and instances on these diagrams. Finding patterns and anomalies in data has important applications in intelligence analysis domains such as crime detection and homeland security. The intelligence community’s focus over many years on improving intelligence collection has come at the cost of improving intelligence analysis. The problem today is often not a lack of information, but instead, information overload. Analysts lack tools to locate the relatively few bits of relevant information and tools to support reasoning over that information. Graph-based algorithms can help intelligence analysts solve this problem by sifting through a large amount of data to find the small subset that is indicative of suspicious or abnormal activity. Till today, large amount of work related to analysis/prediction of threat to national security has been done “manually”. This process is slow and much labor-intensive. Today’s need is tools to analyze large semantic graphs automatically so that we can get the related results in the considerably short span of time. There are many challenges to do these things fast and effectively. The challenges, including how to represent the data effectively, representing temporal information, representing the use of ontology related to them, etc. Other significant challenges include scale and complexity of data and ontologies that are useful for the analysis. We formalize these graphs in this thesis along with provide the efficient way to represent them in logic formalisms. While there are several existing supervised/unsupervised learning frameworks to identify patterns and anomalies from graph data, there has been little work aimed at discovering patterns and abnormal instances in very large semantic graphs whose nodes are richly connected with many different types of links from knowledge representation perspective. To address this problem, we design a novel, disjunctive logic programming framework that utilizes the information provided by different types of nodes and links to identify abnormal nodes and patterns. Our approach represents the dependencies between nodes and paths in the graph in first order logic predicates to capture what we call “semantic profiles” of nodes, and then applies disjunctive logic rules to find abnormal nodes and patterns that are significantly different from their closest neighbors. In a set of experiments on movies data, our system can almost perfectly identify the abnormal instances/patterns