Recent developments in behavior change support applications—such as fitness tracking, mental health monitoring, and smoking cessation—highlight the importance of personalized interventions. In particular, users’ psychological traits have a strong influence on motivation and adherence. However, self-report questionnaires dominate current assessments. These tools require significant time and cognitive effort, often resulting in high attrition rates. Our study presents a model framework that estimates psychological traits from behavioral logs, eliminating the need for questionnaires. We trained the model using questionnaire-derived labels, designed to infer associations between behavioral patterns and underlying psychological traits. This approach supports future low-burden, adaptive interventions. To address the imbalance and bias in psychological trait data, the framework uses a bidirectional data augmentation strategy with Conditional Generative Adversarial Networks (cGANs). This approach enhances data diversity and model robustness. We conducted field experiments with 176 users of a web-based exercise application designed to promote user wellness. Our method surpassed conventional data augmentation techniques, achieving a maximum classification accuracy of 0.98. We also tested generalization using an independent user group and observed an average accuracy of 0.64. These results demonstrate that generative models can facilitate scalable and low-burden psychological trait estimation in real-world settings .