Most of the existing unsupervised feature selection methods tend to obtain an optimal subset of informative features by means of eliminating the noise and reduduncies. Unfortunately, the two kinds of useless features cannot be always removed simultaneously. By reinterpreting the ultimate goal of unsupervised feature selection, we realize that directly selecting useful features can not only naturally avoid both kinds of useless features, but also get a chance to model the interactions between features, which could induce a more explicit interpretation of the result feature subset. To realize the intuition, a half-open concept named the degree of feature cooperation is defined at first and then one implementation of it based on information theory is proposed to quantitatively describe the interaction between features. After that, a framework based on this concept as well as the core idea of hierarchical clustering is further given to select a complementary feature subset as the final output. The experimental result empirically confirms the effectiveness of the proposed method.