Towards A Sequential Decision Model of Interaction with Computer Systems

In this talk I will describe a theory of how people make multi-attribute decisions with visualised data. The theory is based on the assumption that using a visualisation to make a decision is an example of a Partially Observable Markov Decision Problem (POMDP). We illustrate the theory with a model of how a person determines the likelihood of credit card fraud given a number of different situation variables, called cues, including the location of a transaction, its amount, and customer history. Each of these variables may have a different validity and users may weight their relevance to a decision accordingly. The model, which is sensitive to human visual information processing constraints and the task environment, solves the POMDP by learning patterns of eye movements over a visualization of the data. We show that the model does a good job of predicting human performance, and therefore at predicting the value of the visualization.