Data source analysis based on source graph is a common APT detection method. When a host is infected by a Trojan virus, APT attacks enter the attack and intrusion stage, and APT attacks are often mixed in the system behaviors of the infected host in order to obtain host permissions. In this paper, we introduce a self-supervised APT detection method to detect mixed APT attacks in the system behaviors. The method in this paper uses audit logs to build a source graph, and then uses graph representation learning to model benign system entities and behaviors, extract deep features from the source graph, regularize the feature reconstruction part, and then calculate the abnormal scores of system behaviors through anomaly detection methods to find out abnormal system behaviors. This method can perform batch level log APT detection. We evaluated our approach on two widely used data sets. The evaluation results show that good detection results are achieved in these cases, outperforming existing methods in both data sets.