Due to the lack of sufficient interaction support, the recommendation accuracy is poor. To address this, a sparse data collaborative filtering recommendation algorithm based on knowledge graph was proposed. Extract the interaction relationship between users and items, a knowledge graph was constructed, and the entity relationships in the knowledge graph was used to extend the representation of users and items. Combining CNN networks, interactive relationships was expanded into complex structures, contextual information was captured, and similarity using Euclidean distance was calculate. A set of similar neighbors was found for the target user, user collaboration filtering was used to predict ratings, the fusion time weighting strategy was dynamically adjusted, and a recommendation list was generated. Tests have shown that the algorithm has high NDCG values, low MAE and RMSE values, and ideal recommendation performance.