This paper presents a semi-supervised hardware Trojan detection method at the gate level using anomaly detection. In contrast to most supervised learning methods, the proposed method does not need class label information while training, effectively handles the class imbalance problem, and thus is more pragmatic in real-world situations. Furthermore, we ameliorate the existing computation of SCOAP values and propose a novel topology-based location analysis to improve the detection performance. The proposed method outperforms the existing supervised learning methods with fewer features by achieving an overall 99.47% TPR, 99.99% TNR, and 99.99% accuracy.