<p>Attribute reduction based on neighborhood rough set (NRS) is an effective approach for dimensionality reduction and has been widely applied in various fields. Most existing NRS-based attribute reduction methods identify important attributes using classification information in the lower approximation. Only a few studies consider classification information in the upper and lower approximations. However, these methods do not fully capture the complexity of decision consistency information, limiting their ability to effectively exploit valuable information. As a result, they struggle to accurately distinguish the classification abilities of different attribute subsets and lack sensitivity to small differences between them. In addition, these methods use a single fixed neighborhood radius for all attributes, which cannot adapt to the data distributions of different attributes. To address these limitations, a novel attribute reduction algorithm based on multi-neighborhood triple consistency measure (MNTCM) is proposed. First, neighborhoods of varying sizes are generated for each attribute. Next, the consistency between condition and decision attributes is comprehensively examined from three perspectives. On this basis, the multi-neighborhood triple consistency measure is proposed to accurately evaluate attribute importance, and a corresponding attribute reduction algorithm is developed. Experimental results demonstrate that the MNTCM algorithm achieves superior classification performance compared with other advanced reduction algorithms, confirming its effectiveness.</p>