Large-scale annotated data is essential for age prediction in social media, yet obtaining such data is costly. Age is a key psychological and cognitive marker influencing communication and social behavior. Understanding age-related patterns in online interactions can provide insights into cognitive development and identity formation. To address data limitations, we propose a semi-supervised multimodal regression model leveraging Transformer-based variational autoencoders to infer age from textual and social features. This approach aligns with cognitive and social science theories on age-related behavioral patterns. Our framework effectively utilizes unlabeled data, reducing annotation dependency while enhancing predictive accuracy. Empirical results demonstrate superior performance over traditional classification and supervised baselines, advancing interdisciplinary research in age inference and online behavior modeling.