AI Summary & Key Takeaways
AI InsightReapplying prominent approaches on the refined dataset yields clearer insights into inter-app dependencies, highlighting methodological trade-offs and complementary strengths for analyzing complex behavioral data.
📌 3 Key Takeaways
Abstract
Graphical models provide a powerful framework for analyzing complex dependence structures among multiple variables, allowing researchers to uncover direct relationships and reduce spurious correlations in high-dimensional data. This study investigates two prominent approaches—the Gaussian graphical model (GGM) and random graphical model (RGM)—to compare their performance in modeling sparse networks under both simulated and real-world conditions. The GGM method emphasizes computational efficiency through ℓ₁-regularized precision estimation, while RGM adopts a probabilistic view that incorporates uncertainty in graph structures. Through simulation experiments, we evaluate the accuracy, stability, and robustness of each method across varying sample sizes and dimensionalities. A real-world application utilizing smartphone screentime data further illustrates how network-based analyses can reveal behavioral patterns across app usage features, such as time spent and notification frequency. This paper provides the first direct comparison of GGM and RGM on real screentime data, addressing a gap in prior work that considered only pairwise associations without evaluating these models side by side. To enhance data quality, we integrate the Gramian angular field (GAF) technique to convert time-series data into visual similarity matrices, enabling the removal of highly redundant days. Reapplying prominent approaches on the refined dataset yields clearer insights into inter-app dependencies, highlighting methodological trade-offs and complementary strengths for analyzing complex behavioral data.