To address the current limitations of recommendation algorithms in the matchmaking social domain, which often rely on relatively rudimentary data and consequently struggle to achieve higher recommendation accuracy, this study investigates the use of a graph convolutional neural network (GCN) with residual connections for training a recommendation model. The objective is to enhance the predictive probability of successful male-female matches. Initially, we constructed the first dataset in this domain, based on expert knowledge in psychology, featuring a high-dimensional dataset derived from 44 psychological traits and real social relationships. Subsequently, we proposed a GCN model with residual connections based on this dataset and validated the model's predictive performance. Evaluation metrics, including the area under the curve (AUC) and accuracy on the validation set, demonstrated the effectiveness of the training results. The model's validation data indicated a strong performance in predicting male-female matches within social networks, with all positive samples being successfully predicted. By statistically analyzing the high-dimensional feature characteristics of paired nodes in the validation set and comparing them with existing research, we found that the statistical results were consistent with current studies. Therefore, this model holds significant practical value.