VARIABLE COSTS-BASED MULTI-GRANULARITY FEATURE SELECTION WITH TEST COST CONSTRAINT

In recent years, some researchers have studied the cost-sensitive feature selection problem based on the rough set theory. Especially, in view of the variability of test costs and misclassification costs as well as the diversity of feature-value granularities between different features, recently an effective cost-sensitive multi-granularity feature selection approach has been proposed. Nevertheless, the approach does not consider the case where a test cost constraint occurs because of limited resources. To tackle this problem, in this paper a variable costs-based multi-granularity feature selection approach is presented in consideration of test cost constraint. First, based on the theoretic framework called confidence-level-vector-based neighborhood rough set, the test cost-constrained multi-granularity feature selection problem is formally defined. Then a heuristic feature-granularity selection algorithm is designed, by which desirable features and their respective feature-value granularities can be simultaneously selected to minimize the total cost consumed in the test cost-constrained situation. Detailed experiments undertaken on six UCI (University of California - Irvine) datasets validate the effectiveness of the algorithm.