Work

Prior guided multi view clustering characterizes smartphone use behavioral profiles from heterogeneous questionnaire data in adults aged 60 to 69 years

Year: 2026

Type: article

Abstract: Questionnaire surveys often contain heterogeneous modules whose roles in behavioral profiling differ. This study evaluated a prior-guided multi-view clustering procedure using a community survey of smartphone use among 428 adults aged 60–69 years in Chongqing, China. The clustering inputs were organized into a 26-item smartphone-use behavior view, a demographic and sociological view, and a lifestyle and health-habit view; smartphone-use context variables were excluded from cluster construction and examined only after clustering. Candidate solutions with two to six clusters were examined using internal validity and repeated 80% subsampling. The two-cluster solution showed the strongest internal separation. The three-cluster solution had the highest mean subsampling ARI (0.940), retained three nontrivial groups, and was selected as the primary exploratory representation; the five-cluster solution showed similarly high stability but contained a small group of 35 respondents. The three profiles contained 219, 112, and 97 respondents. Conditional on the prespecified prior $\varvec{\pi }=(0.70,0.15,0.15)^\top $ , the estimated view contribution weights were 0.8054, 0.1060, and 0.0886. Behavior-view-only spectral clustering was highly concordant with the full solution (ARI $=0.908$ ; profile correlation $=0.999$ ), showing that the behavior view supplied the core partition while the auxiliary views provided limited, explicitly constrained contextual refinement rather than a qualitatively different structure. Variables used in clustering are reported as profile characteristics rather than independent validation. Smartphone-use duration, timing, location, and function variables that were excluded from clustering showed exploratory post hoc associations with profile membership. The findings support transparent prior-guided integration of heterogeneous questionnaire modules while emphasizing dependence on the specified prior, the dominant behavioral view, data-quality decisions, and the exploratory nature of the profiles.

Source: Scientific Reports

DOI: 10.1038/s41598-026-78680-2