Epithelial–mesenchymal transition (EMT) plays a critical role in promoting cancer metastasis. In this study, cancer EMT is considered as an overall structural change in the gene regulatory network (GRN), and its essential features are elucidated by the network biology approach. We first defined the state space of GRN as a set of all possible activation patterns of GRN, and then introduced the quasi-potential field into this space to show the relative stability distribution of each state. The quasi-potential was determined empirically by collecting gene expression profiles from public databases. Changes of GRN states during the EMT process were traced in the state space, by using time-course data of gene expression profiles of a cell line inducing EMT from the database. It was found that cancer EMT occurred in three sequential stable stages, each of which formed a potential basin along the EMT trajectory. As confirmation, structural changes of GRN were estimated by applying the ARACNe algorithm to the same time-course data, and then applying master regulator analysis to extract the main regulations. Each group of master regulators was found to be alternatively active in the subsequent three stages to cause overall structural changes of GRN during cancer EMT.