In the process of feature selection, information granularity plays an important role in feature representation, which directly determines the accuracy of results. The theory of Granular-Ball Computing enhances model robustness by using coarse-grained objects as input. However, the discarding of certain samples during the process can lead to the loss of critical information, and the coarse-grained structure may result in a loss of detail in feature representation. This paper proposes a neighborhood rough set model that fuses knowledge from two different granular levels through neighborhood relation, effectively addressing these problems. This approach strikes a balance by avoiding information loss while preventing overfitting due to excessive focus on details. We explore the impact of different granularity configurations on the feature selection process. Experiments on public datasets demonstrate that proper granularity configuration can enhance classification accuracy.