MetaEns: Automatic Unsupervised Ensemble Outlier Model Selection

Researchers have introduced MetaEns, an automatic unsupervised framework designed for the selection of ensembles in outlier detection models. By utilizing unsupervised outlier detection, the reliance on labeled data is eliminated, and multi-model ensembles enhance detection reliability. However, forming an ensemble without labeled data poses difficulties, as simple ensembles may experience saturation, where overlapping or unreliable models hinder performance and lead to unnecessary computations. MetaEns leverages labeled meta-datasets to develop a model that forecasts marginal gains from adding a candidate model to an existing ensemble. During testing, this learned signal is integrated with a submodular-inspired proxy objective that promotes diversity-aware discounting and family-level risk regularization, facilitating a greedy sequential selection process. The framework effectively tackles ensemble saturation and aims to autonomously choose efficient outlier detection models without human oversight.