Relief algorithm is a series of feature selection method. It includes the basic principle of Relief algorithm and its later extensions reliefF algotithm. Its core concept is to weight more on features that have essential contributions to classification. Relief algorithm is simple and efficient, thus being widely used. However, algorithm performance is not satisfied when applying the algorithm to noisy and unbalanced datasets. In this paper, based on the Relief algorithm, a feature selection method is proposed, called threshold-Relief algorithm, which eliminates the influence of noisy data on classification results. Combining with the K-means algorithm, two unbalanced datasets feature selection methods are proposed, called K-means-ReliefF algorithm and K-means-relief sampling algorithm, respectively, which can compensate for the poor performance of Relief algorithm in unbalanced datasets. Experiments show the effectiveness of the proposed algorithms.