Viewing stereoscopic images sometimes causes viewers to feel inconvenience, which is called 3D visual discomfort. Excessive horizontal disparity, misalignment between the left and right views, or depth cues conflicts are some of the important factors involved in 3D visual discomfort. The ability to estimate the degree of 3D visual discomfort can be used to improve the 3D display systems and provide acceptable binocular visual quality. Most of the existing visual discomfort prediction (VDP) approaches extract hand-crafted features based on perceptual modeling and statistical analysis of disparities. We have proposed a simple yet effective VDP model based on unsupervised learning of sparse features which are highly predictive of subjective discomfort levels. These features are extracted from the aggregation of left and right disparity maps. This aggregation effectively highlights the areas with sudden changes and high levels of disparities where discomfort is most likely to occur. The regression model trained by the features, predicts high correlated 3D visual discomfort scores on each dataset. The cross-database results are also superior to other reported ones.