Abstract:Anomaly detection is a challenging problem in machine learning, and is made even more so when dealing with instances that are captured in low-level, raw data representations without a well-known and well-behaved set of engineered features. Images or data streams from sensors are good examples of such low-level, raw data representations. The Radial Basis Function Data Descriptor (RBFDD) network is an effective solution for anomaly detection, however, it is a shallow model that does not deal well with low-level, raw data representations. This article investigates approaches to transform an RBFDD network into a deep one-class classifier that works well for anomaly detection problems with low-level, raw data representations. We show that approaches based on simple transfer learning are not effective and our results suggest that this is because the latent representations learned by generic classification models are not suitable for anomaly detection. We show that an alternative approach that adds multiple convolutional layers before the Radial Basis Function (RBF) layer of an RBFDD network—–to form a Deep Radial Basis Function Data Descriptor (D-RBFDD) network—is very effective. This is demonstrated in a set of evaluation experiments based on multiple anomaly detection scenarios created from publicly available image classification datasets, and a real-world anomaly detection dataset in which different types of arrhythmia are detected in electrocardiogram (ECG) data. Our experiments show that the D-RBFDD network out-performs state-of-the-art anomaly detection methods including the Deep Support Vector Data Descriptor (Deep SVDD), One-Class Support Vector Machine (OCSVM), and Isolation Forest on the image datasets, and produces competitive results on the ECG dataset.

Data Classification Based on Density Function Estimated by Radial Basis Function

Novel Data Classification Method Based on Radial Basis Function Networks

Radar: Rare Category Detection Via Computation Of Boundary Degree

Rare Category Detection Algorithm Based on Weighted Boundary Degree

Filling Holes in Point Cloud with Radial Basis Function

A Density-Based Random Forest for Imbalanced Data Classification

Research of Data Mining Approach Based on Radial Basis Function Neural Networks

Training Radial Basis Function Classifiers with Gaussian Kernel Clustering and Fuzzy Decision Technique

A method for mixed data classification base on RBF-ELM network

Health Big Data Classification Using Improved Radial Basis Function Neural Network and Nearest Neighbor Propagation Algorithm

Density-based logistic regression

Design methodology for Radial Basis Function Neural Networks classifier based on locally linear reconstruction and Conditional Fuzzy C-Means clustering

Optimal Bayes Classifiers for Functional Data and Density Ratios

Density Estimation Based on Mass

Relative Density Degree Induced Boundary Detection for One-Class SVM.

Hyperspectral remote sensing image classification based on radical basis function neural network

Image Classification Method Rationally Utilizing Spatial Information of the Image

Discriminant analysis based on kernelized decision boundary for face recognition

An RBF Neural Network Clustering Algorithm Based on K-Nearest Neighbor

The Deep Radial Basis Function Data Descriptor (D-RBFDD) Network: A One-Class Neural Network for Anomaly Detection

An Optimized Neural Network Classification Method Based on Kernel Holistic Learning and Division