<p>In order to solve the classification problem of incomplete hybrid information systems, the concept of neighborhood-tolerance mutual information is defined by combining neighborhood-tolerance relations and mutual information theory in granular computing. And using the idea of ensemble learning, a selection ensemble classification algorithm based on neighborhood-tolerance mutual information (NTMI-SE) is proposed. Firstly, the algorithm obtains information granules according to the missing attributes, divides the granules layer to construct the granules space, and builds a new basic classifier on different granules layers by using the BP neural network ensemble algorithm as the basic classifier. Then, the neighborhood-tolerance mutual information about class attributes was calculated according to the missing attributes of each information granule to measure the importance of each information granule, and the weight of the basic classifier was redefined according to the prediction accuracy of the basic classifier and the neighborhood-tolerance mutual information. Finally, based on the predicted samples, the weighted ensemble prediction results of the basic classifier are analyzed and compared with the traditional ensemble classification algorithm. For partially incomplete hybrid datasets, the proposed ensemble classification algorithm can effectively improve the classification accuracy.</p>