Clinical decision support systems in healthcare practice

Supervised v Unsupervised Learning

In computer science, supervised learning is a technique, usually machine language, that is used in the process of deducing a function from a data set. Typically, the data sets contain pairs of input objects (usually vectors) and desired outputs. The result of this function can be a continuous value, known as regression, or it can predict a classification label of the input objects, known as classification. The task of the supervised learner is both to predict the value of the function for any input that is valid after viewing a universe of training examples. In order to do this, however, the learner must synthesize from the known data (e.g. that which is presented), to hypothetical situations in a way that most individuals would find reasonable (Kotsiantis, 2007). In the field of psychology, these types of tasks are known as concept learning or cognitive attainment (Bruner, et.al., 1967).

In contrast, unsupervised learning is a class of problems in which the major goal is to define the manner in which the data itself is organized. In this paradigm, there are numerous methods used that are often based on data mining methods -- and distinguished from supervised learning in that the learners is given only unlabeled examples. This is quite closely related to the density estimation theory in probability studies in the sense that it is the construction of an estimate based on epistemology (observation of data) focusing on a way to explain probability. This is also often used in population studies and data that requires a subset of materials (e.g. cases of x disease in y population) within a larger population structure. One form of this is called clustering, which allows the research to assign sets of observations into component subsets that have similar properties (Clatworthy, et.al., 2005).

Technically, from a theoretical template, the only real difference between supervised and unsupervised learning is in the causal structure of the paradigm. In supervised learning, the model defines the effect on set of inputs has on the other set of outputs. The inputs are assumed to be at the start of the causal chain, the outputs at the end. The models can, of course, include mediating variable that operate between the input and output sections. In unsupervised learning, all the observations are assumed to be caused by latent variables. For instance, the observations are always assumed to be at the tail end of the causal chains (Petrie, 1996). In actual practice, the paradigms for supervised learning often leave the important probabilities for numerous inputs undefined. If input variables are available, the model works well, but if not, it is difficult to make sense or to infer anything about the outputs (Stork, 2001).

In the lay world, we can think of supervised learning as a computer program that allows an individual to learn a language, say Spanish. The pronunciation of the letters is the output. Each time the right letter is produced; feedback either strengthens or weakens it for a wrong letter. This has about an eighty percent chance of producing correct words. For individuals, feedback is akin to taking a walk and getting poison ivy; exposure to the plants teaches one where to walk, to wear appropriate clothing, and to bring antihistamine cream. In real life situations, too, supervision aids quicker results whereas in unsupervised learning trial and error are the main mode of learning (Goudbeek, et.al., 2006).

REFERENCES and WORKS CONSULTED

Bruner, J., et.al. (1967). A Study of Thinking. New York: Science Editions Press.

Burke, P. (2010). "A Simulation Case Study From an Instructional Design

Framework." Science Direct. March 19, 2010. Cited in:

http://www.sciencedirect.com/science?_ob=ArticleURL&_udi=B82X4-4YN05CG-6&_user=10&_coverDate=04%2F30%2F2010&_rdoc=1&_fmt=high&_orig=search&_sort=d&_docanchor=&view=c&_acct=C000050221&_version=1&_urlVersion=0&_userid=10&md5=759d9102ebe8a27ff567d0db6c90379b