In this paper, we propose a way of effective fraud detection to improve the detection efficiency. We focus on the bias of the training dataset, which is typically caused by the skewed distribution and highly overlapped classes of credit card transaction data and leads to lots of mis-detections. To reduce mis-detections, we take the fraud density of real transaction data as a confidence value and generate the weighted fraud score in the proposed scheme. The effectiveness of our proposed scheme is examined with experimental results on real data.