<p>Feature selection is an important method for reducing the dimensionality of data, effectively removing redundant features to improve the efficiency of an algorithm. To avoid repeated calculations, dynamic feature selection implements a dynamic update strategy when the sample set or feature set changes. Granularity feature selection and incremental feature selection strategies have received considerable attention, but current matrix-based incremental granularity feature selection methods focus on the entire sample set as a means of reducing unnecessary calculations. This paper explores granularity feature selection based on incremental sample changes. First, a relation matrix is derived and used to construct feature (decision) sets. This matrix is modified to obtain a simplified relation matrix. Second, the representative sample set is obtained from the simplified relation matrix and combined with vector inner product calculations. This representative sample-based granularity feature selection algorithm considers four situations that occur when adding new samples. Samples that satisfy certain conditions are classified in terms of positive-region reduction and absolute reduction, and samples that have redundant relationships with the existing representative sample set are deleted. A dynamic update strategy for the discernibility matrix in granularity feature selection is established, and an incremental granularity feature selection algorithm based on representative samples is proposed. Experimental evaluations on public datasets verify the feasibility and effectiveness of the proposed algorithm.</p>