DIMACS Technical Report 2009-12 November 2009 Supporting Cognitive Models of Sensemaking in Analytics Systems

Cognitive science is providing the scienti c community with increasingly well-supported models of the mental stages and representations that professional analysts go through in the course of conducting an investigation, be it reactive or proactive in nature. These process models are generally advanced within the eld of Sensemaking, because the analyst's primary task can be viewed as \making sense" of a large body of unorganized information. One of the most well-known long-running Sensemaking investigations is that of Pirolli and Card et al. [Pir05] Their resulting model provides an initial basis for our research. In using these models to improve analytics systems, we have at least two distinct problems: (1) how to use information about the Sensemaking states gained from user interactions with an analytics system to learn the parameters of an e ective analysis process, and (2) how to use this knowledge to provide user guidance that results in better human-machine interaction and a more robust investigative process. The answers to these questions lie at the intersection of research in machine learning, knowledge representation, user interfaces and cognitive science, and addressing them requires an end-to-end system perspective. In this report, we survey these problems and discuss our approach, system design, and experimental design. In particular, we de ne the Sensemaking model's representation within the software framework, the set of machine learning tasks for learning the parameters of an e cient process, our initial user interface design, the design of the meta-cognitive UI feedback, and nally the design of the initial experiments, including the ground truth which is from an actual solved crime case. We conclude with the insights gained thus far into building interactive systems that support users' cognitive models of Sensemaking.