Recommendation systems, as a kind of information filtering system, can understand users' interests based on their personal data or historical behavior records, and are widely used in Web applications such as e-commerce, search, and streaming websites. Based on the fact that users with similar preferences may be interested in similar items, most existing recommendation methods mine this synergistic information through user-item interactions, called collaborative filtering. Recent interaction information can better reflect users' dynamic interests over time. Another group of work is called temporal-based recommendation, which takes into account temporal information to model the user's dynamic interests. A temporal recommendation system is based on the user's temporal interaction data as a context to predict which item the user is most likely to interact with next time. Markov chain-based recommendation models are a typical example of temporal sequential recommendation, which assume that the next interaction is only related to the previous one. Recently, much work has used recurrent neural networks to model previous behaviors and use hidden states to predict the next behavior. However, almost all temporal recommendation methods model user embeddings based only on their own temporal interaction history, ignoring the heterogeneous information widely available in recommender systems, such as product attributes, and encounter cold-start problems when data are sparse and users have fewer interactions. In this paper we proposed a model, the heterogeneous dynamic graph neural network. This network uses a static graph encoder to process the node representation of each heterogeneous graph at different time steps, which includes the node attribute information and edge information, to capture the changing node information in all time steps of the dynamic graph and obtain a valid node representation at each time step. The long and short-term memory model is then used to aggregate the temporal information between different time steps and to mine the connections and interdependencies of different types of nodes between different time steps. The goal of the heterogeneous dynamic graph neural network is to capture the changing node representations of the same nodes as well as different nodes at different time steps in the graph network, expecting to better perform the task of node classification in dynamic graphs. In this paper, we validate the model of heterogeneous dynamic graph neural networks from data, experimenting with node classification tasks on real graph structured data including social networks, e-commerce networks and business review networks, and show that heterogeneous dynamic graph neural networks outperform the latest representation learning methods for static and dynamic graphs.