Under the background of campus informatization, this paper constructs an intelligent computing framework for college students ‘behavior management. Relying on a university information system, this paper integrates learning platform logs, access control records, dormitory return information, class check-in sequences and campus network access segments to form a multi-source heterogeneous behavior dataset. Through unified time window reconstruction, cross-source semantic mapping and relationship enhancement modeling, the framework combines time convolution, graph attention and early warning decision mechanism to complete behavior recognition, state discrimination and risk stratification. The experimental results show that the behavior recognition accuracy of the proposed method on Camfield-Trace and LMS-Flow datasets reaches 0.931 and 0.918, respectively. The state discrimination accuracy on the synthetic dataset reaches 91.8%, the macro-average F1 value reaches 0.903, and the manual verification agreement rate reaches 93.5%. The average response delay is 84 ms, which indicates that the framework can still maintain a relatively stable output under the condition of complex heterogeneous events, and has the real-time processing ability under the condition of online operation.
