This paper proposes a heterogeneous subgraph transformer (HETEROSGT) for fake news detection. Fake news is prevalent on social media and can cause significant harm to public discourse and societal well-being. The authors construct a heterogeneous graph that captures the relationships among news topics, entities, and content. They find that fake news can be effectively detected through atypical heterogeneous subgraphs centered on them, which encapsulate essential semantics and complex relations between news elements. However, exploring such subgraphs remains challenging due to the heterogeneity of the graph. To address this, the authors propose HETEROSGT, which uses a pre-trained language model to derive both word-level and sentence-level semantics. They then apply random walk with restart (RWR) to extract subgraphs centered on each news, which are fed into a subgraph Transformer to quantify the authenticity. Extensive experiments on five real-world datasets show that HETEROSGT outperforms five baselines in terms of accuracy, macro-precision, macro-recall, macro-F1, and ROC. Further case and ablation studies validate the effectiveness of HETEROSGT's design choices. The key contributions of this work include: (1) the first attempt to explore both word- and sentence-level semantic patterns and structural information among news, entities, and topics for fake news detection; (2) the use of relative positional encoding to mitigate the problem of learning node positional encodings in graph Transformers; and (3) extensive experiments on five real-world datasets demonstrating the superior performance of HETEROSGT over five baselines. The results show that HETEROSGT is effective in detecting fake news by analyzing the irregular subgraph structure and features.This paper proposes a heterogeneous subgraph transformer (HETEROSGT) for fake news detection. Fake news is prevalent on social media and can cause significant harm to public discourse and societal well-being. The authors construct a heterogeneous graph that captures the relationships among news topics, entities, and content. They find that fake news can be effectively detected through atypical heterogeneous subgraphs centered on them, which encapsulate essential semantics and complex relations between news elements. However, exploring such subgraphs remains challenging due to the heterogeneity of the graph. To address this, the authors propose HETEROSGT, which uses a pre-trained language model to derive both word-level and sentence-level semantics. They then apply random walk with restart (RWR) to extract subgraphs centered on each news, which are fed into a subgraph Transformer to quantify the authenticity. Extensive experiments on five real-world datasets show that HETEROSGT outperforms five baselines in terms of accuracy, macro-precision, macro-recall, macro-F1, and ROC. Further case and ablation studies validate the effectiveness of HETEROSGT's design choices. The key contributions of this work include: (1) the first attempt to explore both word- and sentence-level semantic patterns and structural information among news, entities, and topics for fake news detection; (2) the use of relative positional encoding to mitigate the problem of learning node positional encodings in graph Transformers; and (3) extensive experiments on five real-world datasets demonstrating the superior performance of HETEROSGT over five baselines. The results show that HETEROSGT is effective in detecting fake news by analyzing the irregular subgraph structure and features.