Diverse behavior clustering of students on campus with macroscopic attention

Analyzing multi-source heterogeneous behavioral data of individuals in complex environments and discovering effective patterns is a challenging topic. Since cognitive psychology believes that all behaviors can be regarded as attention to different objects, this paper proposes an analysis framework based on Macroscopic Attention (MA) to characterize the diverse behavior of individuals. To verify the effectiveness of the framework, this paper takes the university campus scene as a case study. Driven by online big data from campus networks, WiFi access points, and smart card controllers, MA characteristics, including its stability, span, shifting, and distributivity, are introduced to analyze behavioral patterns. A campus behavior clustering approach based on MA qualities is then proposed to reveal the impact of MA on academic performance, which utilizes a Temporal Convolutional Network (TCN) to extract temporal features. Experiments on behavioral data of over 1,000+ students show that MA-based clustering has advantages in terms of silhouette coefficient, Davies-Bouldin index, and other relevant metrics, compared with clustering based on traditional features. The experiments show that the MA quality distributions of the two student clusters (A and B) follow a Gaussian distribution, but their key parameters show significant statistical differences, such as mean and variance. Additionally, students in Cluster A represent a possibility about 39% higher than students in Cluster B of achieving the top-level scholarship. It is experimentally verified that the stability and distributivity of MA significantly have active impacts on academic performance. The experimental results also show that the stability and distribution of MA are significantly positively correlated with students' academic performance. Based on these discoveries, at-risk students can be detected in time and given advance interventions.