Deep embedding GAN-based model for anomaly detection on high-dimensional sparse data

The use of Generative Adversarial Nets (GAN) for anomaly detection has been explored recently. However, in the case of high-dimensional sparse data, existing GAN-based anomaly detection models suffer from inefficient dimensionality reduction, computationally costly data reconstruction, and suboptimal performance limited by the training objective. In this paper, a deep embedding GAN-based model is developed for anomaly detection on high-dimensional sparse data. In the model, dimensionality reduction of input data is performed by embeddings efficiently. With the bidirectional Wasserstein GAN, data reconstruction is conducted in the input dense representation space at a low computational cost. The objective function defined by the Wasserstein distance and Lipschitz continuity constraints stabilizes training and improves model performance. Experimental results on public datasets show that, the developed model has comparable or superior performance over the competing techniques, and achieves up to 8.81% relative improvement based on the area under the Receiver Operating Characteristics curve (AUC).