In the era of digital socialization, the spread of rumors poses a threat to public safety. Rumor detection therefore becomes critically important. It has been proven that hybrid rumor detection methods can learn richer feature information about rumors and find differences between rumors and non-rumors. However, these methods only learn features of individual rumors and fail to capture the interactive features across multiple rumors’ source posts and comments. To address this issue, we propose a dual perception of source posts relevance and comments flow model DPRD. We construct a heterogeneous information network and design three meta-paths, propose a source posts relevance-perception attention layer to perceive relevance of source posts. Temporal interval strength is applied in the comments flow-perception attention layer to weigh the flowing features of comments. Additionally attribute-level and target-level feature spaces are proposed to prevent the loss of temporal interval attributes. Finally, two features are aggregated through a dual perception fusion layer to obtain the final embeddings. Experiments conducted on three real-world social media datasets demonstrate that the proposed DPRD model outperforms existing baseline models.