<p>To address the problem that the influence of rumor propagation stages is ignored in traditional rumor detection methods, an early rumor detection method based on stage sampling and triple-relationship graph is proposed. Firstly, the sampling probability density function is defined, and the stage sampling method based on normal distribution is proposed to obtain the comment sets with respect to the early stage, middle stage, and late stage of an event. Secondly, both the comment posting time and information propagation directions are considered, and the triple-relationship graph that integrates diffusion relationship, aggregation relationship, and sibling relationship for each stage is constructed. Moreover, the advantage of heterogeneous graph attention network (HAN) in exploring graph structural features is leveraged to obtain the node representations. Finally, to improve the interpretability and to capture the mutual influence between nodes efficiently, a graph-level vector computation method based on compressed self-attention mechanism and soft attention mechanism is proposed. Experimental results on two public datasets show that the proposed method consistently outperforms existing typical methods, with <i>F</i>_<i>w</i> improvements of approximately 2.1% and 3.7% on the CED and Weibo datasets, respectively, validating its effectiveness on early rumor detection. Furthermore, attention weight visualization experiments explicitly highlight the contributions of comments at different stages, significantly enhancing the interpretability of our approach.</p>