Abstract
Local-density anomaly detection remains a core challenge in high-dimensional and large-scale data. The Local Outlier Factor (LOF) is a classic and influential approach, but its dependence on exact nearest neighbor search makes it impractical for modern datasets with hundreds of features or hundreds of thousands of samples. We propose High-Dimensional Local Outlier Factor (HdLOF), a scalable extension of LOF based on a fast and efficient graph-based Approximate Nearest Neighbor (ANN) search, enabling robust density estimation in high-dimensional spaces with dramatically reduced computation time. We further introduce two variants: HdLOF-2L, which leverages a second layer of reachability modeling to capture more complex local structure, and HdLOF-E, an ensemble approach that aggregates multi-scale density information using statistical score fusion. Comprehensive experiments on a suite of real-world tabular, image, and embedding datasets demonstrate that HdLOF and its variants consistently outperform other methods in both accuracy and efficiency for big, high-dimensional data.
Showing the abstract — retrieve the full paper via the Exa API.