Abstract
Social bots increasingly mimic real users and collaborate in large-scale influence campaigns, distorting public perception and making their detection both critical and challenging. Traditional bot detection methods, constrained by single-source features, often fail to capture the complete behavioral and contextual characteristics of social bots, especially their dynamic behavioral evolution and group coordination tactics, resulting in feature incompleteness and reduced detection performance. To address this challenge, we propose CB-MTE, a social bot detection framework based on multi-source heterogeneous feature fusion. CB-MTE adopts a hierarchical architecture: user metadata is used to construct behavioral portraits, deep semantic representations are extracted from textual content via DistilBERT, and community-aware graph embeddings are learned through a combination of random walk and Skip-gram modeling. To mitigate feature redundancy and preserve structural consistency, manifold learning is applied for nonlinear dimensionality reduction, ensuring both local and global topology are maintained. Finally, a CatBoost-based collaborative reasoning mechanism enhances model robustness through ordered target encoding and symmetric tree structures. Experiments on the TwiBot-22 benchmark dataset demonstrate that CB-MTE significantly outperforms mainstream detection models in recognizing dynamic behavioral traits and detecting collaborative bot activities. These results confirm the framework's capability to capture the complete behavioral and contextual characteristics of social bots through multi-source feature integration.