
Boundary Peeling (BP) is a distribution-free framework for unsupervised outlier detection based on successive one-class support vector machine (OCSVM) boundaries. Although BP and its ensemble extension, Ensemble Boundary Peeling (EBP), have shown promising performance, their effectiveness may depend on implementation choices and data characteristics. This study systematically investigates the BP framework’s robustness and tuning through three complementary analyses. First, BP, EBP, and a robustified configuration, Robustified Boundary Peeling (RBP), are evaluated using controlled synthetic datasets covering contamination, dimensionality, correlation, distributional shape, multimodality, and p>n settings, along with semantic benchmark datasets. Second, an ablation study examines the effects of removing BP’s weighting scheme, replacing mean with median aggregation, and adopting a double-MAD-based threshold. Third, the sensitivity of EBP and RBP to stopping rules, OCSVM parameters, ensemble size, and sampled-feature size is assessed. Results show that implementation choices materially affect performance. For EBP, the sampled-feature size matters more than the ensemble size, and a smaller kernel bandwidth improves all average measures. For RBP, the stopping rule and kernel bandwidth affect the thresholded measures, while AUC remains comparatively stable. RBP provides competitive performance, but its advantages are metric- and data-dependent. Overall, the findings emphasize that BP implementation should be treated as a multi-criteria tuning problem rather than a universally optimal fixed configuration.
Fault detection in water treatment (WT) facilities is important to maintain system operation. Batch processes and backwashes within these systems create cyclical data, allowing for the application of functional data techniques. Previously, a limited number of functional fault detection methods have been applied to real data whose faults have unknown start and endpoints. In this article, we use simulations and case studies of cyclical WT data to compare different process monitoring approaches for different types of faults. First, we introduce a new method based on measures of directional outlyingness, and we compare this with three competing methods: (i & ii) Shewhart control charts with and without data preprocessing and (iii) T2 and SPE charts based on functional principal components. A simulation study generating functional data compares the performance of these methods, and then an extensive set of case studies is used to evaluate their real-world behavior. These case studies are unique in that the faults are introduced deliberately, so their start and end times are known. We demonstrate the importance of accounting for expected changing global behavior in WT systems when performing fault detection, as well as the circumstances under which a given fault detection method would be appropriate.
Cross-validation (CV) is a common method to tune machine learning methods and can be used for model selection in regression as well. Because of the structured nature of small, traditional experimental designs, the literature has warned against using CV in their analysis. The striking increase in the use of machine learning, and thus CV, in the analysis of experimental designs, has led us to empirically study the effectiveness of CV compared to other methods of selecting models in designed experiments, including the little bootstrap. We consider both response surface settings where prediction is of primary interest, as well as screening where factor selection is most important. Overall, we provide evidence that the use of leave-one-out cross-validation (LOOCV) in the analysis of small, structured is often useful. More general k-fold CV may also be competitive but its performance is uneven.