Graph Regularized Autoencoder and its Application in Unsupervised Anomaly Detection.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Dimensionality reduction is essential for unsupervised learning tasks like anomaly detection and clustering.
• Autoencoders are commonly used for dimensionality reduction, but effective embedding of nonlinear manifolds requires geodesic distance metrics.
• Existing geodesic distance approximations like ISOMAP inspire the development of new graph-based methods.

Purpose of the Study:

• To propose a novel graph regularized autoencoder that utilizes a minimum spanning tree (MST) for improved dimensionality reduction.
• To enhance the discrimination of data samples by incorporating structure-preserving distances derived from MST.
• To evaluate the effectiveness of the MST-based approach in anomaly detection and clustering applications.

Main Methods:

• A minimum spanning tree (MST) algorithm is employed to approximate local neighborhood structures and generate structure-preserving distances.
• The MST-based distance metric replaces Euclidean distance in the autoencoder's embedding function, creating a graph regularized autoencoder.
• The MST regularizer is integrated into generative adversarial networks (GANs) for further performance evaluation.

Main Results:

• The proposed graph regularized autoencoder significantly outperforms a wide range of alternative methods on 20 benchmark anomaly detection datasets.
• Incorporating the MST regularizer into GANs substantially improves anomaly detection performance.
• The MST regularized autoencoder demonstrates superior performance in clustering tasks on two datasets.

Conclusions:

• The MST-based graph regularization is an effective technique for enhancing dimensionality reduction in autoencoders and GANs.
• This approach significantly improves performance in anomaly detection and clustering tasks.
• The method offers a robust way to preserve data structure in low-dimensional embeddings for high-dimensional data.