Deep learning model has achieved great success in text and image classification tasks. However， it is difficult for a deep learning model to provide an interpretable classification intuition. In this paper， we propose a multi-grained ensemble classification method for unstructured data. Compared with other learning methods， our multi-grained ensemble classification method can preserve the context information of data. In the multi-grained ensemble classification method， data are divided into different granularity， and the data with different granularity are used to train different base learners. Their learning results provide interpretability for the final classification of the ensemble model. Base learners are assigned by different weights according to their performances on the validation set， therefore， a better ensemble learner can be constructed. In the experiments， we verify the validity of our model on three types of unstructured data （i.e.， text， medical images， and time series）， and experimental results show that our model is not only simpler than the state of art but also competitive in the accuracy. Meanwhile， it can provide the interpretability for the final classification results.