Audio Feature Selection Based on Rough Set

Keeping audio features is important for audio index. However, in most cases the features number is huge, thus direct processing is time-consuming. Feature selection, as a preprocessing step of data mining, has turned to be very efficient in reducing dimensionality and removing irrelevant data. In this paper, we propose a feature selection algorithm based on Rough Set theory, which could find out the feature subset from audio stream. The definition of discernibility of ordinal attributes set is introduced to discover the subsets containing implicit features. Moreover, based on the discernibility definition, we consider the discernibility of two and three ordinal attributes set, together with the discernibility of individual attribute, thus the extracted reduct is more complete and meaningful, which is consistent with the experimental evaluations. Keyword: audio feature, Rough Set theory, ordinal attributes set, frame discernibility