One of the successful experiences for e-commerce has been online shopping and its changes to the modern consumer patterns in the market. However, some sellers on shopping websites use strategies to impact evaluation, in order to accumulate credit and enhance sales. This impact evaluation appears to be normal, but becomes a potentially unsafe trading crisis. For example, there are many merchandise sellers with high evaluations that will even sell fakes and make consumers suffer unknowingly. When faced with half-truthed evaluation points, traders not only need to be cautious, but also need to have a faster response with better countermeasures. Because of this, this research found an effective method to detect dishonest sellers on e-commerce sites. First, we collected a huge attribute set to accurately describe the nature of dishonest sellers, covering Rating ,Current ,History and Service of 105 kinds of attributes. Second, we explore the many ways sellers impact evaluations in order to classify the ways sellers impact the evaluations. Next, we use x-means clustering algorithm to separate honest sellers and dishonest sellers. According to the cluster centroids, we can analyze the data and create different detection models. To verify the effectiveness of the proposed method, this present study uses actual transaction data from the Taobao for validation. When using attribute sets of dishonest and honest sellers, the average detection rate is less than 50% accurate. But when using attribute sets of only dishonest sellers, the accuracy rate can be increased to 70% to 78%, indicating the feasibility of the method proposed in this study. When mixing different types of methods of dishonest sellers, accurately detecting dishonest sellers decrease significantly. To conclude, honest and dishonest sellers have very similar methods of selling their merchandise.