Chain Mediation Path Optimization for Smartphone Dependence Prediction Using Graph Neural Networks and Reinforcement Learning

Abstract

Smartphone dependence prediction and subjective well-being impact analysis are challenged by complex variable interactions, suboptimal mediation-chain optimization, and limited interpretability. This paper proposes a chain mediation path optimization framework that formalizes mediation as a directed causal graph, where nodes denote behavioral and psychometric variables and edges encode dependency relations. Graph convolutional neural networks learn high-level node embeddings, and an attention-based multi-scale causal embedding module enhances structural expressiveness at semantic and topological levels. For path optimization, a policy-gradient reinforcement learning module dynamically updates edge weights under a reward that balances prediction accuracy, path consistency, and computational cost. A graph autoencoder with geometric consistency constraints is introduced to stabilize reconstructed mediation chains and preserve structural coherence. The framework is evaluated on 5,724 users, with an average of 58 behavioral features and 12 psychometric indicators per sample, using stratified sampling and cross-validation. The proposed model achieves 92.1%±0.4 Accuracy, 89.4%±0.6 F1-score, and 87.8%±0.5 Topology Score, outperforming a structural equation model baseline and a deep neural network without causal-path constraints. Ablation results verify that attention, reinforcement learning, and geometric constraints jointly improve classification performance and path stability. These results indicate that graph-based mediation path modeling enables unified prediction of dependence and well-being outcomes with improved interpretability and deployability in large-scale settings.