Abstract: Social bots have continuously evolved during their development, posing significant challenges to existing detection models. To address this, we propose a novel social bot detection framework, BotSAI. This framework first employs customized encoders to extract multi-dimensional feature representations from user metadata, text, and heterogeneous social network graphs. Specifically, the graph encoder achieves efficient and balanced aggregation of neighborhood information through oversampling and a local relation transformer. Subsequently, a multi-channel representor maps user representations into invariant subspaces and specific subspaces to enhance their features. Finally, the enhanced user representations are integrated and refined using a self-attention mechanism. Experimental results demonstrate that BotSAI outperforms state-of-the-art methods on two authoritative social bot detection benchmarks. Furthermore, systematic experiments reveal the impact of different social relationships on detection accuracy, providing new research perspectives for social bot detection.