The invention provides an early warning model building method. The method comprises the steps of calling a random training set; and based on the random training set, building an early warning model. Compared with the prior art, the method has the advantages that behavior paths of a large amount of research objects are collected and comparison is performed to generate the training set; a classifier is trained; and the early warning model is generated. Compared with a conventional mode of performing early warning by using a classification algorithm, a device for performing early warning by using the early warning model is more accurate in early warning effect, higher in early warning speed and higher in efficiency, and is suitable to process data with large data and feature quantities.