RGLD Boosts Tabular Anomaly Detection Accuracy and Speed

Unsupervised tabular anomaly detection requires methods that are not only accurate and robust across various datasets but also computationally efficient. Traditional statistical detectors often lack flexibility, while deep learning methods, though powerful, are typically slow and difficult to tune without supervision. This research introduces RGLD (Randomized Global-Local Density Estimation), a novel approach designed to address these challenges. RGLD combines two complementary branches: a global random-feature density branch that identifies samples in broadly low-density regions, and a local neighbor branch that detects samples poorly supported by nearby observations. Both branches operate over feature-bagged randomized views of the data, allowing RGLD to uncover anomaly evidence that might be obscured in a single data representation. Extensive experiments on 47 tabular datasets, comparing RGLD against 23 statistical and deep anomaly detection baselines, demonstrated its superior performance. RGLD achieved the strongest dataset-level AUROC, ranking first in dataset wins, and second in AUPRC wins. Crucially, it also proved significantly faster than all evaluated deep detectors, achieving 50x-580x speedups, while remaining competitive with statistical methods in runtime, offering an excellent accuracy-efficiency trade-off.