CLOE Enhances Anomaly Detection in High-Dimensional Data.

Anomaly detection is crucial across many sectors, but existing lightweight methods often struggle with complex, high-dimensional datasets and require extensive hyperparameter tuning. While Christoffel Function-based methods are appealing for their simplicity and strong theoretical basis, their scalability to high-dimensional data has been a significant limitation. A new method, CLOE (Christoffel Loss Autoencoder), addresses this by integrating an autoencoder for efficient dimensionality reduction with a Christoffel Function-based detector operating in the compressed latent space. CLOE introduces an innovative loss function that leverages the Christoffel Function to optimize the autoencoder's representation learning, ensuring it better captures the underlying distribution of normal data. The approach also provides a principled way to set detection thresholds and efficiently tune its single hyperparameter. Experimental results on various high-dimensional tabular anomaly detection benchmarks show CLOE's superior performance while maintaining the simplicity and low-tuning advantages of Christoffel Function-based techniques.