STEM education is a hot issue in modern education, and it is important to study whether middle school students enter STEM careers in the future in the early stage of career planning. In this paper, we collected students’ behavioral data through the online tutoring platform ASSISTments, divided the raw log data into five types: single-valued, binary-valued, multi-valued, continuous-valued and cumulative, and aggregated them using different data reconstruction methods. Then, a width & depth prediction model based on feature crossover is proposed to perform feature crossover on the aggregated data, and then the depth and width models are jointly trained using. During the training process, the AUC of the FC-Wide&Deep model improved rapidly from 0.800 to 0.845 in the 1st to 16th training rounds, and then slowly climbed with the increase of training rounds. By averaging the results of the three tests, the AUC index of the FC-Wide&Deep model test results improved by 1.29% compared to the DNN model, and the RMSE index improved by 2.08% compared to the BSN-FM model. The FC-Wide&Deep model is generalizable and generalizable, and can be applied to predict whether students will enter STEM careers in the future, thus contributing to the cultivation and leadership of STEM talents in the field of education.