Reproduce “Generative Adversarial Active Learning for Unsupervised Outlier Detection” on OpenResearch

About this paper

Yezheng Liu, Zhe Li, Chong Zhou, Yuanchun Jiang, Jianshan Sun, Meng Wang, Xiangnan He

Multiple-Objective Generative Adversarial Active Learning (MO-GAAL) is proposed for unsupervised outlier detection, utilizing a GAN with multiple sub-generators to produce diverse and informative potential outliers. This approach allows a discriminator to define robust decision boundaries, improving detection accuracy and scalability compared to several existing methods, especially on complex, high-dimensional datasets.

Key insights

• Generative adversarial networks can be directly utilized to generate 'hard negative samples' or informative potential outliers, transforming unsupervised outlier detection into a robust binary classification task.
• The issue of mode collapse in GANs can be effectively addressed by employing multiple sub-generators, each tasked with learning distinct modes or clusters of the normal data distribution.
• Intelligent generation of potential outliers can circumvent the limitations of traditional unsupervised methods, particularly in high-dimensional spaces plagued by data sparsity and the curse of dimensionality.

Summary by alphaXiv.