| Summary: | Any observation that follows a pattern other than the expected one, i.e., the normal behaviour, is considered abnormal behaviour (also known as an anomaly). Abnormal behaviour is witnessed in various areas— for instance, a previously unseen high temperature during winter in a naturally cold environment.
Prior research shows that anomalies can result in negative impacts such as financial losses in telecommunications or human fatalities in aviation accidents. Despite the advances made in the area of anomaly detection, detection methods underperform due to the challenges that affect and hinder the process of anomaly detection. In this work, three novel anomaly detection approaches are introduced each, of which aims at addressing one problem that debilitates the performance of anomaly detection.
One of the problems in anomaly detection is having access to an ample amount of anomalous examples; therefore, the proposed methods in this work are all unsupervised as this type of learning is needless of having access to a labelled training set. The first contribution of this research is focused on reducing the execution time of a density-based method while maintaining the performance at a high level by applying a novel pruning-based preprocessing step.
In density-based methods, measuring the density plays an important role, and as the dimensionality increases, the definition of density becomes harder. By using dimensionality reduction methods, it is possible to transform the high-dimensional input data into a low-dimensional form while maintaining essential features. In the second contribution, a novel dimensionality reduction method is introduced that is needless of having access to an anomaly and noise-free training set.
When using One-Class Classifier methods, the performance varies as the size of the training set changes. Having access to a training set that includes more normal examples can improve the performance as the class-boundary becomes less ambiguous. The final contribution of this work is focused on improving the definition of class-boundary by proposing a data augmentation approach. The proposed approach generates augmented examples while simultaneously reduces the dimensionality of the input data.
|
|---|