The categories presented are neither meant to be exhaustive nor mutually exclusive: Predictive analytics, for example, utilizes system models and is usually used for either automated or non-automated decision making. Likewise, decision making utilizes both a model of a system and, most likely, predictions derived from this model regarding the values of unknown variables in the system.
Rather these categories, and the placement of the associated examples, are meant to assist those familiar with terminology from various sectors understand the relation of Bayesian Network technologies to the categorical frameworks they are used to.
Examples: Scientific Systems Modeling, Social Systems Modeling
A system, or variables thereof, is identified to be of interest. A model, or meta-model, is generated from (potentially incomplete) datasets collected from the system and/or via working with experts to model domain knowledge.
The network model generated by this process will contain chance nodes, each representing a particular variable of the system, and potentially utility nodes, indicated the utility of specific states the system can be in. The network generated includes easily identifiable representations of the conditional independencies found between variables of the system, and probability distributions representing the correlations present. These provide a uniquely powerful tool for understanding of the relationships present in the system modeled, which can then be employed to predict potential states of the system and to make decisions regarding the optimum method of manipulating the system. They also indicate the sets of variables that, if known, provide complete knowledge of the system's influence on given variables of interest, making future data collection easier and more cost-efficient.
For example, take a organization that wishes to model the relationship between infrastructure, social practices and the standards of health within a particular region. The organization has access to a large amount of data regarding variables in all three categories at given time point and utilizes these to learn a network model of the 'system' in question. The model learnt provides an understanding of how these variables interact. In particular it indicates whether certain variables are conditionally independent of others, how the variables interact in their effects on one another, and the extent of the correlations present.
Examples: Portfolio Analysis, Marketing Analysis, Sector/Economy Analysis, Client Analysis, Risk Analysis, Fraud Detection and Prevention, Geostatistics, Security Analysis.
Once a system has been modeled, certain variables of interest which are difficult to independently ascertain are selected. These variables are represented by prediction nodes, and, given a data vector of known values for some subset of the variables of the model, the model utilizes the discovered conditional independencies and correlations to predict the probability of the variables of interest taking particular values.
Notice that while this process includes classification, in that the model will be able to indicate which value a variable is most likely to take in a given situation, it entails much more than this. Unlike, say, neural networks and their associated models, the Bayesian Network techniques employed provide a probability estimation for the variable taking any of the possible values. It can also provide error bar estimations of the confidence of these probability estimates!
For example, take a client who is interested in predicting the movement in the dollar price of oil. A set of fifty variables is identified that domain experts believe to be importantly related to this variable of interest. Historical data is used to model the conditional independencies and correlations present between these variables. Current conditions (the current values of these fifty variables) are then presented to the model, which used them to predict the likely movement of the oil price in dollars. The model not only indicates which ranged of movement is most likely, and gives its best estimate of this likelihood, but also provides probability estimates for the movement being in any given range, and calculates the bounds within which it is 99.5% certain the actual probabilities lie.
Decision Support, Decision Automation and System Optimization
Decision Automation and Support Examples: First-Line Management Automation, Medical Diagnostics.
System Optimization Examples: Manufacturing, Logistics, Human Resources, Industrial Systems, Education, Agriculture, Environmental Protection.
Just as a system model can be used for predictions, it can also be used for automated decision making. In this case, variables which the client is able to control are identified and are designated decision variables. Also identified are variables which are not currently known but will be so prior to such decisions being assigned a value (i.e. prior to the decision being made). These are termed information priors and can be other decision variables.
The Inatas System Modeler and Inatas System Viewer software packages are then able to run algorithms which determine the best decision given the specified utilities and the potential values the relevant information priors take. This produces what is known as a policy for each decision, which permits either automated or simple manual look for the best decision given what is known of the current state of the system. Note that this includes not only the value of known variables but the predicted values of all unknown variables given the state of the known variables!
The client can then decide what should be done with the generated decision policies. Automation of response is a useful option for first-line decision procedures such as whether to flag a product as potentially defective, to alter variables within certain limits in an automated system in response to environmental conditions or, in conjunction with a live feed of data input from the system, as a means of continual system optimization.
Alternatively such policies can be used as a decision support tool, potentially replacing a layer of human management. For example, a staff member may, when faced with an issue which they are unsure how to handle, in the first instance consult such a policy rather than a human manager. Only if the results continue to be unsatisfactory would they proceed to request managerial assistance.
Such decision automation and support is obviously helpful in, for example, the management of industrial or logistic systems. Note, though, that it is also useful in the management of human systems. A company struggling to retain skilled employees, or a school-system seeking to help underachieving students, can use such techniques to provide insightful hypotheses about how to alter the variables under their control in the given problem's context in order to maximize the outcomes they wish to achieve.