With the development of the Internet, educational data based on social networks and mobile applications are also emerging, and personalized teaching systems have become a topic that has gradually attracted attention. In this paper, after exploring the meta-learning algorithm based on deep learning, a dynamic graph neural network is used to predict user preference, and a recommendation algorithm model is constructed based on meta-learning. On this basis, combined with the teaching and training mechanism of CDIO, a teaching system for the piano improvisation accompaniment course of preschool education was built, and the performance of the construction system was tested to test its application effect in the piano improvisation accompaniment course of preschool education. It can be seen that the values of each index of the recommended algorithm model in this paper are between 0.382~0.956 under cold and warm start, and the effect is the best. And the MRR@10 accuracy of the model in this paper is up to 0.725. In the professional piano improvisation accompaniment course, the scores of all indicators of the system use effect in this paper are above 4 points, which are higher than those of the traditional system, which proves that the method can improve the use effect of the teaching system and provides practical value for the system design of professional piano improvisation accompaniment course.