The cognitive diagnostic assessment and constructed digital safety literacy assessment framework is based on the following five aspects: Information Identification (IJ), Self-Protection (SP), Safe Communication (SC), Legal Compliance (LC) and Safety Emergency Response (SE). With the help of the seq-GDINA multi-level scoring model, an empirical investigation about college students was made. Multidimensional improvement paths include educational intervention measures, cooperation between school and family, community and teacher, policy guarantee and other methods. Diagnostic results show that only 13.50% (42 people) can grasp all attributes well. Among all the students, nearly half have the deficiency in LC and SE attributes. And SC is in the worst condition, whose low-score part mastery probability is as low as 20.29%. Comparison of modeling shows that seq-GDINA is superior to the conventional model and the neural network model with advantages in ACC (0.811), RMSE (0.377) and F1 score (0.818). Experimental teaching also verifies the effectiveness of the improvement path, and more than 85% of the students think their level has been improved in general.
