The key to numerous attribute selection methods lies in the utilization of a given measure as the attribute evaluation criterion, along with the constraints of heuristic algorithms. However, the absence of attribute similarity evaluation and the simplistic sequential selection mechanism result in redundant attribute traversal, leading to significant time consumption. Additionally, the use of a single measure limits the perspective of attribute evaluation, making it difficult to unearth attributes with high learning performance. In view of this, a framework for attribute selection is proposed, where: 1) Attribute grouping is performed based on attribute granularity and knowledge distance between attributes. Within each group, the attributes exhibit significant differences, while between groups, the attributes possess strong discriminative power. This allows attribute traversal to be conducted at the group level, effectively compressing the search space of candidate attributes and improving attribute selection efficiency. 2) The proposed restricted Pareto optimality principle is utilized to iteratively select attribute groups, ultimately obtaining the desired subset of attributes. In experiments conducted on 12 UCI datasets by injecting four different levels of attribute noise, the results show that compared to 8 popular methods, the proposed approach yields attribute selection results with an average improvement of 5.89% in classification stability, an average improvement of 12.28% in classification accuracy, and an average reduction of 59.27% in time consumption.