A decision support method, based on bounded rationality concepts, to reveal feature saliency in clustering problems

In many real-life data mining problems, there is no a-priori classification (no target attribute that is known in advance). The lack of a target attribute (target column/class label) makes the division process into a set of groups very difficult to define and construct. The end user needs to exert considerable effort to interpret the results of diverse algorithms because there is no pre-defined reliable “benchmark”. To overcome this drawback the current paper proposes a methodology based on bounded-rationality theory. It implements an S-shaped function as a saliency measure to represent the end user's logic to determine the features that characterize each potential group. The methodology is demonstrated on three well-known datasets from the UCI machine-learning repository. The grouping uses cluster analysis algorithms, since clustering techniques do not need a target attribute.

► A method for data mining problems with no a-priori classification ► S-shaped function is used for feature selection via feature saliency detection. ► The S-shaped function is based on bounded rationality concepts. ► The S-shaped function pinpoints the feature values that characterize each cluster. ► The method is illustrated on three real life datasets using clustering algorithms.