This paper addresses the task of friend recommendation in the VKontakte social network based on a multimodal analysis of user profiles. We utilize textual data, photographs, music preferences, friend lists, group memberships, and the social interaction graph of 52 volunteers. For each data modality, embeddings are extracted: textual (via BERT), visual (via ResNet), audio (via a pre-trained audio embedding model), and graph-based (via GCN/GraphSAGE). These embeddings are integrated into a unified multimodal model. Experiments are conducted across four evaluation scenarios: real existing friendships, user-perceived relevance, oracle-based profile similarity, and cross-cluster recommendations. We report performance metrics including Precision@5, Recall@5, F1, Accuracy, NDCG, and MAP for each scenario. Results show that the comprehensive multimodal model consistently outperforms graph-only and content-only approaches in recommendation accuracy. The advantages of the multimodal strategy and its contribution to intelligent technologies in social engineering are discussed.