Hierarchical Text Classification (HTC) is a complex task that involves categorizing text within a hierarchical labeling system. Current methodologies primarily focus on integrating hierarchical information into encoding models to achieve enhanced text representations. However, the intricate nature of the process obtaining representation often diminishes the model’s interpretability, leaving researchers uncertain about the model’s effective utilization of hierarchical information and the rationale behind its results. In this paper, we introduce a novel Hierarchical Text Classification method based on Selection-Inference(HTCSI) using large language models (LLMs) to enhance interpretability. Specifically, we leverage classification paths within the labeling hierarchy as cue information to guide LLM reasoning. Additionally, we develop a path selection module to identify the most probable classification paths, thereby minimizing the impact of irrelevant information on LLM reasoning. We employ the Parameter-Efficient Fine-Tuning (PEFT) method to bolster the reasoning capabilities of LLMs, thus improving the accuracy of hierarchical text classification while ensuring a more reliable reasoning process. Experiments conducted on several standard datasets demonstrate that HTCSI not only enhances credible classification reasoning but also surpasses other advanced HTC methods.