CLOE: New anomaly detector beats benchmarks with one hyperparameter

CLOE (Christoffel Loss Autoencoder) is a new semi-supervised anomaly detection method proposed by Léa Billet, Louise Travé-Massuyès, Elodie Chanthery (LAAS, INSA Toulouse, ANITI), and Alexandre Gaffet. It addresses the scalability limitations of Christoffel function–based detectors, which are theoretically elegant and require at most one hyperparameter but struggle with high-dimensional data. CLOE resolves this by first using an autoencoder to reduce dimensionality, then applying a Christoffel function detector in the latent space. Crucially, the researchers introduce a custom loss function that leverages the Christoffel function to push the autoencoder toward representations that better capture the support of the normal data distribution, aligning representation learning directly with the anomaly detection goal.

The method also includes a principled procedure for setting the detection threshold and an efficient strategy for tuning its single remaining hyperparameter. Experiments across multiple high-dimensional tabular anomaly detection benchmarks show that CLOE outperforms existing methods while preserving the simplicity and low-tuning advantages of Christoffel function approaches. The work was presented at the 16th International Conference on Information Science and Technology (ICUIS 2026) in Coimbra, Portugal. For practitioners who need robust, lightweight anomaly detection in fields like process monitoring, healthcare, or finance, CLOE offers a compelling balance of performance and ease of use.