Global and Local Structure Discrimination for Effective and Robust Outlier Detection

Deep outlier detection on high dimensional data is an important research problem with critical applications in many areas. Though promising performance has been demonstrated, we observe that existing methods characterized outliers only from a single perspective, which leads to reducing the distinction of inliers/outliers with the growth of training epochs. This in turn hurts the robustness and effectiveness of outlier detection since the optimal training epoch on a special dataset is unknown in unsupervised scenarios. In this paper, we propose a DNN based framework with both global and local structure discrimination for effective and robust Outlier Detection, named GOOD. The global module compacts the data (mainly inliers) since the majority of data are inliers, while the local module scatters the data (mainly outliers) based on that outliers reside in low-probability density areas. These two modules are cleverly united by a self-adaptive weighting strategy that trades off the degree of complementary and competitive cooperation. The complementary views can help effectively detect outliers with diverse characteristics, and such competitive learning can prevent a single module from learning the entire data too well and ensure robust detection performance. Comprehensive experimental studies on datasets from diverse domains show that GOOD significantly outperforms state-of-the-art methods by up to 30 \(\%\) improvement of AUC while performing much more robustly with the growth of training epochs.