Deep Neural Network Prediction of Couple Relationship Quality Using Multidimensional Relational and Psychological Features

Abstract

Objective: This study aimed to develop and evaluate a deep neural network model for predicting couple relationship quality using multidimensional relational, psychological, contextual, and demographic features.

Methods and Materials: This quantitative cross-sectional predictive study was conducted with 680 adults involved in committed romantic relationships and residing in Canada. Participants completed standardized assessments of relationship quality, relationship satisfaction, communication quality, conflict processes, perceived partner responsiveness, emotional intimacy, commitment, trust, attachment anxiety and avoidance, depressive symptoms, anxiety, perceived stress, emotion regulation, self-esteem, psychological well-being, and selected contextual characteristics. The dataset was divided into training (n = 476), validation (n = 102), and independent test (n = 102) subsets. A feedforward deep neural network with four hidden layers was trained using the Adam optimizer, dropout regularization, batch normalization, and early stopping. Predictive performance was evaluated using mean absolute error, mean squared error, root mean squared error, and coefficient of determination and was compared with multiple linear regression, ridge regression, random forest regression, and gradient-boosted regression. SHAP analysis was used to interpret feature importance.

Findings: The deep neural network demonstrated the strongest predictive performance on the independent test set, with MAE = 3.91, MSE = 26.63, RMSE = 5.16, and R² = .84. It outperformed multiple linear regression (R² = .64), ridge regression (R² = .66), random forest regression (R² = .74), and gradient-boosted regression (R² = .78). The most influential predictors were relationship satisfaction, perceived partner responsiveness, destructive conflict, communication quality, emotional intimacy, attachment anxiety, commitment, trust, perceived stress, and attachment avoidance. Sensitivity analysis showed that removing relationship satisfaction reduced model performance to R² = .76, while the multidimensional full model retained the highest accuracy.

Conclusion: Deep neural networks can accurately predict couple relationship quality by integrating relational and psychological features and may provide a useful framework for identifying complex nonlinear patterns relevant to relationship assessment and intervention.