\underline{R}everse \underline{e}ngineered class \underline{d}iagrams (REDs) are helpful to ease the comprehension of complex software. However, the original REDs might contain many details and thus provide little benefit. Condensing REDs by identifying the most important classes (aka \textit{key classes}) and discarding unimportant ones has been regarded as a promising way. In the last decade, many \underline{k}ey \underline{c}lass \underline{p}rediction (KCP) approaches have been proposed. However, the \textit{unweighted network metrics} used in these studies fail to capture the \textit{dependency strength} between classes and thus cannot precisely measure the actual complexity of classes. In this paper, we propose an approach, KEEPER, for condensing REDs, which introduces a set of weighted network metrics to characterize the complexity of classes and to build KCP models. Empirical studies on a set of 18 open-source Java projects show that i) KEEPER performs better than all the baseline approaches; ii) different weighting schemes do not impact the intra-scheme relative effectiveness of KEEPER variants, but they do affect the inter-scheme relative effectiveness of KEEPER variants; and iii) KEEPER has a good scalability and thus can be used in practice efficiently.