Numerous applications, such as bank transactions, road traffic, and news feeds, generate temporal datasets, in which data evolves continuously. To understand the temporal behavior and characteristics of the dataset and its elements, we need effective tools that can capture evolution of the objects. In this paper, we propose a novel and important problem in evolution behavior discovery. Given a series of snapshots of a temporal dataset, each of which consists of evolving communities, our goal is to find objects which evolve in a dramatically different way compared with the other community members. We define such objects as community trend outliers. It is a challenging problem as evolutionary patterns are hidden deeply in noisy evolving datasets and thus it is difficult to distinguish anomalous objects from normal ones. We propose an effective two-step procedure to detect community trend outliers. We first model the normal evolutionary behavior of communities across time using soft patterns discovered from the dataset. In the second step, we propose effective measures to evaluate chances of an object deviating from the normal evolutionary patterns. Experimental results on both synthetic and real datasets show that the proposed approach is highly effective in discovering interesting community trend outliers.