It takes a network to want a child: Explaining fertility intentions with graph neural networks

Abstract

How does one’s social network influence their fertility intentions, and which network variables are most influential? In this thesis the application of graph neural networks is explored to analyse how social networks influence fertility intentions. We trained and applied a heterogeneous neural network model using transformer convolutional layers. To enhance model interpretability, integrated gradients and Shapley value sampling explainer techniques were employed, leading to different attribution scores for input features. The model achieved an average test R2 score of 0.4129 across 50 random data splits. This indicates that the model captures relevant patterns although significant variability in the data remains unexplained. The explainer technique integrated gradients emphasized the importance of alter and certain ego variables, whereas Shapley value sampling predominantly highlighted ego features. These discrepancies showcase the critical role of an explainer’s underlying mechanism in interpreting the model’s outputs. Despite the complexities that come with an individual’s fertility decision, this study demonstrates that graph neural networks can reveal meaningful relationships within personal networks, and highlight which variables are most important in predicting fertility intentions. The challenge of generalizing predictions for distinct individuals persists, as fertility intentions fundamentally reflect personal choices. Future work could enhance model robustness by targeting specific subgroups within larger, more diverse datasets. Though the difficulties of explaining such models remain a challenge, graph neural networks can uncover the patterns that influence life-altering decisions, including those related to fertility.