Dual-MGAN: An Efficient Approach for Semi-supervised Outlier Detection with Few Identified Anomalies