Feature selection is frequently used in data pre-processing for data mining. When the training data set is too large, sampling is commonly used to overcome the difficulty. This work investigates the applicability of active sampling in feature selection in a filter model setting. Our objective is to partition data by taking advantage of class information so as to achieve the same or better performance for feature selection with fewer but more relevant instances than random sampling. Two versions of active feature selection that employ class information are proposed and empirically evaluated. In comparison with random sampling, we conduct extensive experiments with benchmark data sets, and analyze reasons why class-based active feature selection works in the way it does. The results will help us deal with large data sets and provide ideas to scale up other feature selection algorithms.