UnDBot is an unsupervised, interpretable, and effective framework for detecting social bots based on structural information theory. The framework constructs a multi-relational graph to model social user behaviors, utilizing three social relationship metrics: Posting Type Distribution, Posting Influence, and Follow-to-follower Ratio. These metrics capture various aspects of social bot behaviors and are used to build a weighted social multi-relational graph. A novel method for optimizing heterogeneous structural entropy is introduced, which aggregates edge information from the graph to generate a two-dimensional encoding tree. This tree enables hierarchical clustering of social bots. A new community labeling method is also proposed, which distinguishes social bot communities by computing user stationary distribution, measuring user contributions to network structure, and counting user aggregation intensity within the community. Comprehensive experiments on four real social network datasets demonstrate that UnDBot outperforms existing social bot detection approaches in terms of effectiveness and interpretability. The framework is designed to be practical, effective, and interpretable, making it suitable for real-world social bot detection tasks. The main contributions of this work include the proposal of an unsupervised and interpretable social bot detection framework, the development of a new social multi-relational graph, the introduction of a new heterogeneous structural entropy optimization method, and the development of a new community labeling method. The framework is evaluated on four datasets, showing its effectiveness, interpretability, and efficiency in detecting social bots.UnDBot is an unsupervised, interpretable, and effective framework for detecting social bots based on structural information theory. The framework constructs a multi-relational graph to model social user behaviors, utilizing three social relationship metrics: Posting Type Distribution, Posting Influence, and Follow-to-follower Ratio. These metrics capture various aspects of social bot behaviors and are used to build a weighted social multi-relational graph. A novel method for optimizing heterogeneous structural entropy is introduced, which aggregates edge information from the graph to generate a two-dimensional encoding tree. This tree enables hierarchical clustering of social bots. A new community labeling method is also proposed, which distinguishes social bot communities by computing user stationary distribution, measuring user contributions to network structure, and counting user aggregation intensity within the community. Comprehensive experiments on four real social network datasets demonstrate that UnDBot outperforms existing social bot detection approaches in terms of effectiveness and interpretability. The framework is designed to be practical, effective, and interpretable, making it suitable for real-world social bot detection tasks. The main contributions of this work include the proposal of an unsupervised and interpretable social bot detection framework, the development of a new social multi-relational graph, the introduction of a new heterogeneous structural entropy optimization method, and the development of a new community labeling method. The framework is evaluated on four datasets, showing its effectiveness, interpretability, and efficiency in detecting social bots.