Scalable Outlier Detection Using Distance Projections

Data exploration is a critical component in data science. However, this has become increasingly more difficult for a large, high-dimensional dataset that contains dirty records. In this paper, we propose a solution for detecting outliers in large high-dimensional datasets using distance-based projections. Our proposed solution first maps the high-dimensional data to multiple one dimension values using their distances to a random set of reference points. For each projected value of a data point, we compute a local outlier factor score and then obtain a combined score that can be used to detect outliers. Our solution is computationally much cheaper than traditional methods. We demonstrate the effectiveness of our solutions using both simulation and real data studies.