fuzzy relation partial fuzzy logiccomposition feature selectorclassificationdragonfly

Compositions of partial fuzzy relations dealing with undefined values have been extensively studied in partial fuzzy set theory. Their effectiveness in practical classification problems has been addressed. In general, the compositions relate to three sets of objects, truth-valued features and classes. The aim of compositions is to assign a class to an object by using knowledge of the assignment of features to objects and classes to features. In a classical setting, all features of objects (subjects to the classification) are known in advance. Such an assumption may be indeed inconvenient for some practical applications, e.g., because of the costly determination of features. This paper aims at the development of such framework that allows obtaining results from the composition even if some features are unknown and, more importantly, to select the most appropriate feature to be questioned additionally to maximize the accuracy of the result. For this purpose, the concept of dynamic feature selectors is defined. Based on classification results obtained from the composition on mandatory features, a dynamic feature selector recognizes the most promising optional feature whose knowledge may best improve the accuracy of classification. We propose two distinct dynamic feature selectors: (1) based on the sharpness and maximal truth values of the result, the ability to reach maximal values, and (2) based on the distance between the original and updated result of the composition. A practical experiment shows the performance of the proposed solution.