Gradual patterns highlight correlations between attributes through rules such as "the more/ less X, the more/less Y" where X and Y are features. These patterns represent valuable knowledge for experts. In the literature, numerous methods allow their extraction based on a binary representation introduced in the GRITE algorithm. Although some of these methods allow for parallel processing, they consume significant memory and require substantial execution time. In this paper, we present a criterion that reduces the number of candidate patterns and, consequently, their execution time. Through experiments conducted on real and synthetic data, we compared the effect of the proposed criterion on the performance of the GRITE and Paraminer algorithms. The results show a significant improvement.