Decision support algorithms form an important part of the larger world of data mining. The purpose of a decision support system is to provide a human user with the context surrounding a complex decision to be based on computational analysis of the data at hand. Typically, the data to be considered cannot be adequately managed by a human decision maker because of its volume, complexity or both; data mining techniques are therefore used to discover patterns in the data and inform the user of their saliency in terms of a particular decision to be made. Visualization plays an important role in decision support, as it is through visualization that we can most easily comprehend complex data relationships (Tufte, 1997, 2001, 2006; Wright, 1997). Visualization provides a means of interfacing computationally discovered patterns with the strong pattern recognition system of the human brain. As designers of visualization for decision support systems, our task is to present computational data in ways that make intuitive sense based on our knowledge of the brain’s aptitudes and visual processing preferences. Confidence, in the context of a decision support system, is an estimate of the value a user should place in the suggestion made by the system. System reliability is the measure of overall accuracy; confidence is an estimate of the accuracy of the suggestion currently being presented. The idea of an associated confidence or certainty value in decision support systems has been incorporated in systems as early as MYCIN (Shortliffe, 1976; Buchanan & Shortliffe, 1984).
TopBackground
A decision support system functions by taking a set of rules and evaluating the most preferable course of action. The most preferable of a set of possible actions is chosen based on an internal optimization of some form of objective function. This optimization may take one of several forms: a full cost-benefit analysis (Rajabi, Kilgour & Hipel, 1998; Hipel & Ben-Haim, 1999); a simple best-rule match; or that of a multi-rule evaluation using rules weighted by their expected contribution to decision accuracy (Hamilton-Wright, Stashuk & Tizhoosh, 2007).
The underlying rules forming the structure of a decision support system may be found using an automated rule discovery system, allowing a measure of the quality of the pattern to be produced through the analysis generating the patterns themselves (Becker, 1968); in other cases (such as rules produced through interview with experts), a measure of the quality of the patterns must be made based on separate study (Rajabi, Kilgour & Hipel, 1998; Kononenko & Bratko, 1999; Kukar, 2003; Gurov, 2004a,b).
The construction of a tool that will assist in choosing a course of action for human concerns demands a study of the confidence that may be placed in the accurate evaluation of each possible course. Many of the suggestions made by a decision-support system will have a high-risk potential (Aven, 2003; Crouhy, Galai & Mark, 2003; Friend & Hickling, 2005). Examples of such systems include those intended for clinical use through diagnostic inference (Shortliffe, 1976; Buchanan & Shortliffe, 1984; Berner, 1988; de Graaf, van den Eijkel, Vullings & de Mol, 1997; Innocent, 2000; Coiera, 2003; Colombet, Dart, Leneveut, Zunino, Ménard & Chatellier, 2003; Montani, Magni, Bellazzi, Larizza, Roudari & Carson, 2003; Devadoss, Pan & Singh, 2005) and medical informatics (Bennett, Casebeer, Kristofco & Collins, 2005): other systems may have a lower immediate risk factor, but the long term public risk may be extensive, such as in environmental planning and negotiation (Rajabi, Kilgour & Hipel, 1998; Freyfogle, 2003; Randolph, 2004).