The cultivation of responsibility consciousness in colleges and universities has entered the scene where course learning, collective collaboration, social practice, network behavior and daily education and management work together. Traditional evaluation relies on questionnaires, conversations and stage summaries, which can reflect students’ explicit attitudes, but it is difficult to consistently identify the formation process of responsible behaviors. In this paper, we construct a reviewable responsibility awareness data package, build a comprehensive index of responsibility around 1,248 anonymous students, 21 behavioral characteristics, 9,984 weekly intervention records, and 35,944 heterogeneous relational edges, and propose a contextual graphical attention responsibility recognition model. The model incorporates students, course tasks, practical activities, collaborative projects, and educational feedback into the same heterogeneous graph, and identifies student responsibility states through relational attention, temporal smoothing, fairness constraints, and explanatory outputs. The results show that the model in this paper has an AUC of 0.906, an Accuracy of 0.823, a Macro-F1 of 0.831, and an RMSE of 0.294 in responsibility state identification, which are 0.060, 0.062, 0.069 and 0.067 higher and lower than the XGBoost, respectively.The results of the 8-week intervention show that the composite index of responsibility taking of the synergistic intervention group increased from 63.8 to 74.6, the proportion of low concern decreased from 28.4% to 15.9%, and the sustained participation rate increased from 39.4% to 57.6%. The ablation experiment showed that the social practice mapping, reflective text features and temporal attention module contributed the most to the model performance. The study illustrates that the digital cultivation of responsibility awareness needs to shift from single-behavioral statistics to multi-source contextual modeling, and to form a closed loop of educational management through interpretive feedback, hierarchical interventions, and manual review.
