This study proposes CIFW, a framework for personalized generation and adaptation of civic education content that integrates self-organizing mapping network SOM and deep learning, aiming to address the limitations of traditional recommendation methods in data sparsity, dynamic adaptation of user interests, and matching of resources across education stages. The low-dimensional mapping and dynamic clustering of user behaviors and item features are achieved by introducing an improved Item-SOM model, combining a multilayer perceptron MLP with bilinear feature interaction technology. The CIFW model is further proposed to optimize the feature weight allocation by using the channel attention mechanism and to enhance the higher-order feature combination capability by bilinear interaction. The experiment is based on the data of 30 users’ ratings of six types of Civic Education resources, and comparing the MAE value and coverage rate, it is found that the MAE value of Item-SOM-CIFW is 0.754 and the coverage rate of 68.9% significantly outperforms that of the traditional algorithms User-CF and FCMCF. The test of the number of matches by grades reveals that the fitness value of this model at different stages of the college freshman to the senior year, e.g., G5’s 1469 group improves up to 43.3% compared with the control group, which verifies the adaptability of dynamic recommendation for Civic Education.
