<p>The rapid growth of social networks and advancements in artificial intelligence have led to a surge in social bots within online spaces, raising concerns about information authenticity and cybersecurity risks. Current methods for detecting social bots can be broadly classified into machine learning, deep learning, and graph-based approaches. Although graph-based techniques are commonly used, they primarily depend on explicit relationships to construct graphs, limiting their ability to uncover implicit connections between users and fully capture the varying strengths of these relationships. To address these limitations, this paper introduces BotICC, a novel detection framework that combines Variational Autoencoder (VAE) technology to infer implicit connections from user data. These inferred connections are integrated with explicit relationships to enhance the graph structure. Additionally, BotICC employs a hierarchical pooling gated network to better capture relationship influence strength, improving the accuracy of social bot detection. Our experimental results show that BotICC outperforms the latest baseline models, achieving accuracy improvements of 0.35% and 0.91% on two benchmark datasets.</p>