Robust isolation forest using soft sparse random projection and valley emphasis method

Isolation Forest (iForest) is an unsupervised anomaly detection algorithm that identifies anomalies based on the assumption that they are “few and different.” While many studies have attempted to enhance iForest, the resulting algorithms often exhibit substantial performance variation across datasets. Importantly, prior improvements in split strategies have frequently overlooked the underlying data distribution, making them ineffective at isolating sparse and widely distributed anomalies, even in relatively simple cases such as unimodal feature distributions. To address this limitation, we propose Robust iForest (RiForest)—a novel variant designed to enhance both performance and stability. RiForest integrates original variables with random hyperplanes generated via soft sparse random projection to select more effective split features in a dataset-independent manner. It also employs the valley emphasis method, a relatively underexplored technique, to determine optimal split points. Furthermore, RiForest introduces sparsity randomization into the soft sparse random projection process, improving robustness against noise. Experiments on 24 benchmark datasets demonstrate that RiForest consistently outperforms existing algorithms in anomaly detection tasks, underscoring its stability and resilience to noisy variables.