Distance Based Joint Probability Density Estimation for Unsupervised Outlier Detection

Outlier detection is a vital preprocessing step in data mining and it holds a great importance for Machine Learning (ML) algorithms. If a ML model is learned without removing the outliers from the data, the outliers present in the data can influence the prediction accuracy of a ML model and the outcome of such a model can be misleading. Keeping in view the importance of outliers detection, this paper proposes an unsupervised outlier detection mechanism. The proposed outlier detection mechanism is based on the Joint Probability Density Estimation (JPDE) with an integration of a Distance Measure (DM). The proposed approach has an advantage of utilizing only a single dimensional distance vector to compute the outliers in a dataset. This enables the proposed algorithm to find the outliers from a high dimensional dataset with low computational complexity. Furthermore, three different approaches based on JPDE-DM are proposed and evaluated using some complex benchmark synthetic datasets.