Temporal link prediction (TLP) aims to predict future links and is attracting increasing attention. The diverse interaction patterns and nonlinear nature of temporal networks make it challenging to design high-accuracy general prediction algorithms. Black-box models such as network embeddings and graph neural networks have gradually become the mainstream for TLP, mainly due to their high prediction accuracy. However, a good TLP algorithm also needs to assist us in exploring the network evolution mechanism. Accuracy-oriented black-box methods cannot sufficiently explain the evolution mechanism because of their low interpretability. Hence there is a need for a high-accuracy white-box TLP method. In this paper, we turn the perspective of link prediction to node itself, a more microscopic level whose dynamic nature we take to predict future links. Two dynamic properties – node activity and node loyalty – are extracted and quantified. Activity is the basic ability of a node to obtain links, and loyalty is its ability to maintain its current link state. Based on the above two properties, we propose a Develop-Maintain Activity Backbone (DMAB) model as our TLP algorithm. Comparative experiments with six state-of-the-art black-box methods on 12 real networks illustrate that DMAB has excellent prediction performance and well captures network evolution mechanisms.