Network data is ubiquitous in real-life applications, such as social media, advertising recommendation and intelligent decision making. Network representation learning techniques map elements in the network onto a low-dimensional vector space, and have found widespread application in downstream tasks such as node classification. In recent years, node2vec methods in the field of network representation learning have been used to generate low-dimensional representation vectors of network nodes by collecting network structural features through random walk and using the word2vec language model. However, the two hyperparameters controlling breadth-first search and depth-first search in node2vec are given globally and uniformly, which makes the random walker insufficiently adaptive. In the network, different nodes have different structural characteristics, and the more specific hyperparameters need to be set according to the structural characteristics of the nodes. The multi-level tree structure mapping method proposed in this paper first maps nodes to different levels by calculating the importance of nodes. Different depth and breadth search intensities are adopted for network nodes at different levels, allowing the random walker to adopt different search strategies at different nodes. Thus the algorithm captures the structural features more effectively. Through experiments on realistic network data, it can be seen that the proposed multi-level tree mapping (MTMap) method can effectively improve the performance of the node2vec method.