Pre-training Graph Neural Networks to Predict Women’s Fertility Intentions from Ego-Network Surveys

Abstract

Understanding fertility intentions—the expressed likelihood of having a (further) child—is essential for addressing demographic challenges linked to low birth rates. While prior research has emphasized individual-level predictors, social networks play a critical but often overlooked role in shaping reproductive decisions. Due to nuanced nature of social interactions, training AI models for tasks like fertility intention prediction typically requires large amounts of labeled data. However, gathering sufficient manually labeled data can be time-consuming and costly, especially when dealing with smaller datasets, such as the personal network data available for this study. This research addresses this challenge by investigating whether Graph Neural Networks (GNNs), enhanced via self-supervised pre-training tasks like masked attribute reconstruction, can capture network dynamics to improve predictions of women’s fertility intentions, particularly given the constraints of smaller graph datasets. Using personal network data from the Dutch LISS panel, we construct ego-alter graphs and evaluate the effects of pre-training strategies on downstream regression performance. While the reconstruction task effectively reduces pre-training loss, it offers no consistent benefit for fertility intention prediction compared to models without pre-training. Ablation studies further reveal that masking ratios strongly influence training dynamics but do not substantially affect final accuracy. These findings suggest that while GNNs offer a promising approach for representing social context, careful alignment between model objectives, graph structure, and target behavior remains essential, especially when using self-supervised techniques to overcome data limitations.