A0201
Title: Non-decimated lifting based outlier detection algorithm
Authors: Nebahat Bozkus - Giresun University (Turkey) [presenting]
Abstract: Outlier detection techniques typically generate outlier scores, after which the researcher must establish a threshold to distinguish between inliers and outliers. A novel approach is introduced that assigns empirical probabilities of being outliers to individual objects on a dendrogram using the non-decimated lifting algorithm. The proposed algorithm first removes noise from the hierarchically built tree using the non-decimated lifting transform, then assigns a probability of being an outlier to each object on the tree. Subsequently, the algorithm eliminates objects with high probabilities from the tree and assigns an empirical probability of being a cluster to each node on the tree. This approach is called Non-Decimated Lifting-based Outlier Detection (NDLout). The performance of NDLout is compared with other existing approaches in the literature using real-world and synthetic datasets.
