Abstract. Time-series classification is a widely examined data mining task with various scientific and industrial applications. Recent research in this domain has shown that the simple nearest-neighbor classifier using Dynamic Time Warping (DTW) as distance measure performs exceptionally well, in most cases outperforming more advanced classification algorithms. Instance selection is a commonly applied approach for improving efficiency of nearest-neighbor classifier with respect to classification time. This approach reduces the size of the training set by selecting the best representative instances and use only them during classification of new instances. In this paper, we introduce a novel instance selection method that exploits the hubness phenomenon in time-series data, which states that some few instances tend to be much more frequently nearest neighbors compared to the remaining instances. Based on hubness, we propose a framework for score-based instance selection, which is combined with a principled approach of selecting instances that optimize the coverage of training data. We discuss the theoretical considerations of casting the instance selection problem as a graph-coverage problem and analyze the resulting complexity. We experimentally compare the proposed method, denoted as INSIGHT, against FastAWARD, a state-of-the-art instance selection method for time series. Our results indicate substantial improvements in terms of classification accuracy and drastic reduction (orders of magnitude) in execution times.