Many deep learning algorithms have achieved satisfactory results on many supervised learning tasks, but they rely on a large number of labeled samples, and the classifiers trained with specific categories can only classify these categories. Zero-shot learning wishes that the computer can reason like a human, it uses historical knowledge to infer the characteristics of new objects and has the ability to recognize novel categories without lots of samples. It is found that there are sparse matrix and "cold-start" phenomena in zero-shot learning task, these phenomena are also in the recommendation tasks. Inspired by the recommendation tasks, the zero-shot classification task is modeled as a matrix completion problem, hoping to learn from the collaborative filtering algorithms in the recommendation field, which regards the sparse labeled matrix as the product of the visual feature matrix and semantic feature matrix, and then classifies the novel samples. In order to make the semantic representation of each category more accurate, a semantic graph structure is constructed based on the semantic relations between categories and a graph neural network is applied on it for information transferring between known and novel categories. Traditional zero-shot learning and generalized zero-shot learning experiments are performed on three classic zero-shot learning data sets. The experimental results show that the collaborative filtering based zero-shot learning method proposed in this study can effectively improve the classification accuracy, and the training cost is relatively small.