Survey on extreme learning machines for outlier detection

In a two-class classification task, if the number of examples of one class (majority) is much greater than that of another class (minority), then the classification is said to be class imbalanced. It can occur among many real-world applications, such as intrusion detection, medical diagnosis, etc. The class imbalance issue can make learning difficult since learning opts to bias towards the majority class. Outliers are cases with anomalous behaviors and are extreme cases of class imbalance. Despite late advances in extreme learning machines (ELMs), there are not many experimental investigations in the field of ELM with outlier detection. In this survey, we provide a comprehensive overview of existing ELMs to address the problem of outlier detection under a unified perspective. Firstly, we describe the background of our work, which includes a brief overview of previous surveys and a detailed description of the enhanced ELMs. Next, available studies regarding why ELMs are used to tackle the class imbalance problem are reviewed. Furthermore, cutting-edge algorithms are surveyed for improved ELMs to detect outliers. We classify these methods under three different machine learning perspectives (i.e., supervised, unsupervised, and semi-supervised approaches). In addition, we explore the developments of existing solutions based on three standardized quality metrics (i.e., accuracy, robustness, and speed) and other performance metrics (e.g., mean absolute percentage error and mean absolute error). After that, related datasets are detailed to facilitate future studies in this field. Last but the most important, this study concludes with discussions, challenges, and suggestions to guide future research.