A hybrid dimensionality reduction method for outlier detection in high-dimensional data

Outlier detection becomes challenging when data are featured by high-dimension. Using dimensionality reduction (DR) techniques to discard the irrelevant attributes is a straightforward solution. However, it appears to be rather difficult for single DR algorithm to discover all outliers, owing to the rarity, heterogeneity, and boundless nature of outliers. In this paper, we propose a hybrid DR method dedicated to outlier detection base on ensemble learning. Multiple algorithms with different specifications of parameters are used to generate accurate and diverse base detectors at the phase of ensemble generation. A two-stage combination function is used at the phase of ensemble combination. Both variance reduction and bias reduction are taken into account in our framework. More importantly, the high flexibility of the proposed detection framework implies that any outlier detection algorithm can be applicable. 15 high-dimensional data sets from KEEL repository and one image data set are used to validate the performance of our method. One semi-supervised and one unsupervised outlier detection algorithms are used in separate experiments. In spite of subtle differences, the advantage of our method has been approved by both experiments. Moreover, contributions of two ingredients of our method are also verified via two pairs of experimental comparisons.