Particle swarm optimisation for outlier detection
Ammar W. Mohemmed, Mengjie Zhang, Will Neil Browne · 2010
Outlier detection is an important problem as the underlying data points often contain crucial information, but identifying such points has multiple difficulties, e.g. noisy data, imprecise boundaries and lack of training examples. In this novel approach, the outlier detection problem is converted into an optimisation problem. A Particle Swarm Optimisation (PSO) based approach to outlier detection can then be applied, which expands the scope of PSO and enables new insights into outlier detection. Namely, PSO is used to automatically optimise the key distance measures instead of manually setting the distance parameters via trial and error, which is inefficient and often ineffective. The novel PSO approach is examined and compared with a commonly used detection method, Local Outlier Factor (LOF), on five real data sets. The results show that the new PSO method significantly outperforms the LOF methods for correctly detecting the outliers on the majority of the datasets and that the new PSO method is more efficient than the LOF method on the datasets tested.