Extended Isolation Forest

We present an extension to the model-free anomaly detection algorithm, Isolation Forest. This extension, named Extended Forest (EIF), resolves issues with assignment of score given data points. motivate problem using heat maps for scores. These suffer from artifacts generated by criteria branching operation binary tree. explain this in detail and demonstrate mechanism which it occurs visually. then propose two different approaches improving situation. First we transforming randomly before creation each tree, results averaging out bias. Second, is preferred way, allow slicing use hyperplanes random slopes. approach remedying artifact seen maps. show that robustness algorithm much improved method looking at variance scores points distributed along constant level sets. report AUROC AUPRC our synthetic datasets, real-world benchmark datasets. find no appreciable difference rate convergence nor computation time between standard EIF.