Outlier Detection Techniques using Sparse Grids Density Estimation

Outlier detection helps to improve results of a clustering process by identifying noisy, anomalous data points in a dataset.However, lots of techniques for outlier detection require a density estimation of the data points, which cannot be computed exactly.To deal with this problem, spatially adaptive sparse grids can be used to learn and approximate the underlying density function of a multi-dimensional dataset.After this learning process, also known as sparse grids density estimation, the obtained approximated function can be evaluated at every data point to receive a corresponding density value.In this thesis, several outlier detection techniques including the Local Outlier Factor and the Local Density Factor are presented.Furthermore, a new densitybased approach to obtain a factor for the outlierness of a data point is introduced.The purpose of this thesis is to assess whether outlier detection is a suitable field of application for sparse grids density estimation.To this end, this approach is integrated into an outlier detection framework allowing comparison to other known methods.To validate the results of the presented outlier detection techniques, several artificial datasets with a certain percentage of outliers are tested.Additionally, real datasets are used for further expirements and analysis of the studied detection methods.iii