Generative adversarial nets for unsupervised outlier detection

Outlier detection, also known as anomaly detection, has been a persistent and active research area for decades due to its wide range of applications in various fields. Many well-established methods have difficulty fitting the distribution of high-dimensional and complex data, making it difficult to detect outliers that have a low degree of deviation. To address this problem, we combine the distribution fitting capability of generative adversarial nets (GANs) with the specificity of the outlier detection problem and propose a GAN-based unsupervised outlier detection (GUOD) method. In a real dataset mixed with normal objects and outliers, the generator of GANs prefers to fit the distribution of the majority of normal objects to minimize the error; as a result, the generated fake data can be used as an augmentation of normal objects. Next, fake “normal objects” are used to train the autoencoder. Finally, the real data are fed into the autoencoder for one forward propagation, and the reconstruction error of the object is used as its own outlier factor. The top-n objects with the largest reconstruction errors are considered outliers. Extensive experiments are conducted on eight real-world datasets, and the results show that the GUOD method performs better than ten other state-of-the-art algorithms.