Petermann, A.

On Pattern Mining in Graph Data to Support Decision-Making

Dissertation, Univ. Leipzig

2019

Dissertation

Futher information: https://nbn-resolving.org/urn:nbn:de:bsz:15-qucosa2-331662

Abstract

In recent years graph data models became increasingly important in both research and industry. Their core is a generic data structure of things (vertices) and connections among those things (edges). Rich graph models such as the property graph model promise an extraordinary analytical power because relationships can be evaluated without knowledge about a domain-specific database schema. This dissertation studies the usage of graph models for data integration and data mining of business data. Although a typical company's business data implicitly describes a graph it is usually stored in multiple relational databases. Therefore, we propose the first semi-automated approach to transform data from multiple relational databases into a single graph whose vertices represent domain objects and whose edges represent their mutual relationships. This transformation is the base of our conceptual framework BIIIG (Business Intelligence with Integrated Instance Graphs). We further proposed a graph-based approach to data integration. The process is executed after the transformation. In established data mining approaches interrelated input data is mostly represented by tuples of measure values and dimension values. In the context of graphs these values must be attached to the graph structure and aggregated measure values are graph attributes. Since the latter was not supported by any existing model, we proposed the use of collections of property graphs. They act as data structure of the novel Extended Property Graph Model (EPGM). The model supports vertices and edges that may appear in different graphs as well as graph properties. Further on, we proposed some operators that benefit from this data structure, for example, graph-based aggregation of measure values. A primitive operation of graph pattern mining is frequent subgraph mining (FSM). However, existing algorithms provided no support for directed multigraphs. We extended the popular gSpan algorithm to overcome this limitation. Some patterns might not be frequent while their generalizations are. Generalized graph patterns can be mined by attaching vertices to taxonomies. We proposed a novel approach to Generalized Multidimensional Frequent Subgraph Mining (GM-FSM), in particular the first solution to generalized FSM that supports not only directed multigraphs but also multiple dimensional taxonomies. In scenarios that compare patterns of different categories, e.g., fraud or not, FSM is not sufficient since pattern frequencies may differ by category. Further on, determining all pattern frequencies without frequency pruning is not an option due to the computational complexity of FSM. Thus, we developed an FSM extension to extract patterns that are characteristic for a specific category according to a user-defined interestingness function called Characteristic Subgraph Mining (CSM). Parts of this work were done in the context of GRADOOP, a framework for distributed graph analytics. To make the primitive operation of frequent subgraph mining available to this framework, we developed Distributed In-Memory gSpan (DIMSpan), a frequent subgraph miner that is tailored to the characteristics of shared-nothing clusters and distributed dataflow systems. Finally, the results of use case evaluations in cooperation with a large scale enterprise will be presented. This includes a report of practical experiences gained in implementation and application of the proposed algorithms.