Abstract
Subjective Well-Being (SWB) is a fundamental metric for assessing the mental health and quality of life of middle-aged and elderly populations. Traditional machine learning models often suffer from low complexity and poor universality, leading to inaccurate predictions by treating individuals as independent samples. This paper introduces DL-GNN, a multi-relational graph neural network designed to capture the complex structural information and shared characteristics among individuals. By modeling social and demographic relations as a graph, the proposed method achieves a 90.25% F1 score on the CHARLS dataset, significantly outperforming traditional baselines.
Introduction
As China's population ages—with nearly 20% of citizens now aged 60 or older—predicting and understanding Subjective Well-Being (SWB) has become a critical public health priority. Traditional predictive models, such as Multiple Linear Regression or standard Multi-Layer Perceptrons, typically rely on the "individual feature hypothesis," which assumes that each survey respondent is an independent sample. This approach fails to account for the structural relationships and shared psychological patterns between individuals with similar life circumstances.
To address these limitations, we propose the use of Graph Neural Networks (GNNs) to model the non-linear relationships within socio-psychological data. By treating individuals as nodes in a multi-relational graph, we can leverage the context of "relational neighbors" to improve prediction accuracy.
Methods
The DL-GNN model architecture operates through a three-stage pipeline:
- Relational Graph Construction: Nodes (individuals) are connected based on four specific relations: shared age, shared positive factors, shared negative factors, and shared life satisfaction levels.
- Similarity Neighbor Node Selector: To mitigate noise and prevent the "over-smoothing" common in GNNs, the model utilizes a Fully-Connected Network (FCN) to calculate similarity between a central node and its neighbors. Using top-p sampling, the model adaptively filters out low-similarity neighbors.
- Neighbor Relation Aggregator: A dual-layer attention mechanism is employed. Intra-relation aggregation weighs the importance of neighbors within a single relation, while inter-relation aggregation weighs the importance of the relations themselves.
The final objective function combines prediction loss, similarity calculation loss, and $L_2$ regularization to ensure robust learning.
Datasets and Experimental Setup
The study utilizes the China Health and Retirement Longitudinal Study (CHARLS) dataset (2011–2018), comprising 17,092 respondents aged 45 and older. Each respondent is characterized by 78 features, including demographic, family, and health variables. The target label is a binary classification of happiness derived from the CESD-10 depression scale and life satisfaction surveys.
Results
The DL-GNN model was compared against several baseline methods, including traditional machine learning (XGBoost) and standard graph-based models (GCN, SDCN).
As shown in the results, DL-GNN achieved superior performance across all metrics, with an F1 score of 90.25% and an AUC of 93.31%. This represents a significant improvement over standard MLPs (F1: 76.14%) and standard GCNs (F1: 81.54%). Notably, experimental analysis revealed that a single-layer GNN architecture performed best for this dataset, as increasing depth to three layers led to overfitting.
Conclusion
The DL-GNN model demonstrates that incorporating relational context is vital for accurately predicting the subjective well-being of elderly populations. By filtering irrelevant connections and weighing multiple social relations, the model provides a more nuanced understanding of happiness. Future work will explore dynamic graph structures to capture temporal changes in well-being and focus on improving the interpretability of the model's decision-making process.
Summary updated