Recommendation system is becoming increasingly important in various fields of life. To guarantee the accuracy of recommendation, the detection of shilling attacks must be considered. However, the performance of the existing detection techniques for shilling attacks is relatively low, especially for unknown types of attacks, the existing detection techniques are not universal. In this paper, we propose an improved clustering algorithm-based shilling attacks detection method. This method uses information entropy to select a feature and uses the selected feature to calculate the similarity between two users in the clustering algorithm. Experiments show that the algorithm has good detection performance in the detection of shilling attacks.