The paper presents UniDBot, an unsupervised, interpretable, and effective framework for detecting social bots. It leverages structural information theory to model social user behaviors and construct a unified, weighted social multi-relational graph. The framework includes three key components: multi-relational graph construction, user community division, and community binary classification. Multi-relational graph construction captures various aspects of social bot behaviors, such as posting type distribution, posting influence, and follow-to-follower ratio. User community division uses a novel method to optimize heterogeneous structural entropy, generating a two-dimensional encoding tree for hierarchical clustering. Community binary classification combines community influence (measured by stationary distribution) and cohesion (measured by node entropy) to distinguish social bot communities. Extensive experiments on four real datasets demonstrate the effectiveness and interpretability of UniDBot, showing superior performance compared to existing models. The main contributions include an interpretable framework, a novel multi-relational graph, an optimized structural entropy method, and a community labeling technique.The paper presents UniDBot, an unsupervised, interpretable, and effective framework for detecting social bots. It leverages structural information theory to model social user behaviors and construct a unified, weighted social multi-relational graph. The framework includes three key components: multi-relational graph construction, user community division, and community binary classification. Multi-relational graph construction captures various aspects of social bot behaviors, such as posting type distribution, posting influence, and follow-to-follower ratio. User community division uses a novel method to optimize heterogeneous structural entropy, generating a two-dimensional encoding tree for hierarchical clustering. Community binary classification combines community influence (measured by stationary distribution) and cohesion (measured by node entropy) to distinguish social bot communities. Extensive experiments on four real datasets demonstrate the effectiveness and interpretability of UniDBot, showing superior performance compared to existing models. The main contributions include an interpretable framework, a novel multi-relational graph, an optimized structural entropy method, and a community labeling technique.