TDE: Trimmed Density Estimation and Its Application in OutlierDetection; An Approach Based on Forward Search

The presence of an outlier in data influences the accuracy and reliability of statistical inference. Density estimation is affected by an outlier too. We propose Trimmed Density Estimation (TDE) approach which by forward search in density of observations, ranks the data based on their outlyingness. This technique is able to detect a local outlier as good as a global outlier in multivariate mixture data. The introduced algorithm, after trimming the outliers, estimates the density of all observations using the remaining clean data to achieve robust density estimation. TDE overcomes the limitations of existing methods, such as the low accuracy and high sensitivity to initial setting of the parameters. To implement the forward search, we present a technique to select an initial free outlier subset in complex data. Using the estimated density of ordered data by forward search, we figure out a cut-off point to trim the outliers. Experiments demonstrate that TDE’s average accuracy is better than the existing approaches and is free of tuning the parameters.