An Application of Heterogeneous Agents to Fabricate Large, Realistic Corporate Transaction Data Sets for Data Mining Tool Testing and Evaluation

We describe methods used to specify and instantiate hundreds of heterogeneous agents and their use in a simulation of national trade and shipping. The agents, representing synthetic corporate entities, interacted to produce hundreds of thousands of trade transaction documents. The goal for the system was for the corpus of documents to evidence diverse, but realistic linkage patterns of corporate entities engaged in emergent shipping behaviors. We then used the documents to test and evaluate a data mining tool that purported to be able to detect these types of behavioral patterns. Our contributions include a design algorithm for a heterogeneous MAS that produces multi-featured outcomes, and a method for instantiating realistic heritages and preferences of agents that extends recent work in heterogeneous random utility modeling.