Abstract: Outlier detection aims to identify data that significantly deviates from the majority of samples in order to reveal unusual patterns or rare phenomena in the data. It holds significant applications in fields such as finance, healthcare, cybersecurity, and industry. However, most deep learning-based outlier detection methods primarily focus on the feature extraction capability of neural networks for individual samples while overlooking the potential relationships among data samples. Moreover, in high-dimensional and sparse feature datasets, the performance of unsupervised methods is easily affected by irrelevant features and noise, making it difficult to distinguish outliers with subtle differences from normal samples, thereby limiting their generalization ability. To address the above issues, this paper proposes a Self-supervised Subgraph Sampling for Outlier Detection (SSGOD). SSGOD constructs an adaptive nearest neighbor graph to fully capture the similarity relationships among tabular data and employs a graph masking strategy to train the autoencoder, guiding the encoder to effectively learn general feature representations and distribution information. Based on this, this paper designs an anomaly subgraph sampler that can adaptively select anomaly subgraphs more relevant to outlier patterns and perform outlier detection through a trained encoder and fully connected classification network. SSGOD has been extensively experimented on 15 real-public datasets, and the results show significant advantages over 11 state-of-the-art outlier detection methods.