In this paper, we propose a novel algorithm called active learning-based nearest neighbor mean distance (ALNNMD) novelty detection method. And this method can be applied to huge collections of data. ALNNMD is based on the framework of active learning. In each iteration it can choose the instance that most optimizes the current novelty detection model from the data pool, and then remove this instance to the training set. Thus, this method can just use few data samples to get a better novelty detection model, having the advantage of both reducing the size of data sets and optimizing the novelty detection model. Furthermore, a comparative experiment is carried out on the UCI data sets, pointing out the effectiveness of the proposed method.