Unsupervised Outlier Detection Using Memory and Contrastive Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Outlier detection aims to identify anomalous data points within a dataset.
• Current deep learning methods often use reconstruction tasks, assuming outliers are harder to reconstruct.
• Auto-encoder (AE) based models may fail to reliably detect outliers due to unconstrained feature learning.

Purpose of the Study:

• To propose a novel unsupervised outlier detection method.
• To address limitations of reconstruction-based deep learning approaches for outlier detection.
• To enhance outlier detection by focusing on feature space analysis.

Main Methods:

• Developed a Memory and Contrastive learning based Outlier Detection (MCOD) method.
• Utilized a memory module to enforce feature consistency for normal data (inliers).
• Employed a contrastive learning module to learn discriminative features, separating inliers from outliers.

Main Results:

• MCOD demonstrated strong performance on four benchmark datasets.
• The proposed method successfully outperformed eleven existing state-of-the-art outlier detection techniques.
• Experiments validated the effectiveness of feature space analysis for unsupervised outlier detection.

Conclusions:

• Unsupervised outlier detection can be effectively performed in the feature space.
• The MCOD method, combining memory and contrastive learning, offers a robust approach.
• MCOD shows significant potential for improving anomaly detection systems.