Motor imagery brain–computer interfaces are one of the widely adopted techniques for imparting basic communication capability to motor disabled patients. The preciseness of a motor imagery BCI task classification is highly dependent on identifying the subject-specific relevant subset of frequency filters. This article proposes a novel approach that utilizes graph theory-based unsupervised feature selection method to determine a reduced set of non-redundant and relevant frequency bands. The empirical analysis of the proposed method is conducted on publicly available datasets, and the obtained results show improvement in classification performance. Further, the performed Friedman statistical test also establishes that the proposed approach surpasses the baseline techniques in classification accuracy. © 2021, The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.