Unsupervised Anomaly Detection through Mass Repulsing Optimal Transport
Authors: Eduardo Fernandes Montesuma, EL HABAZI Adel, Fred Maurice NGOLE MBOULA
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through a series of experiments in existing benchmarks, and fault detection problems, we show that our algorithm improves over existing methods. Our code is publicly available at https: //github.com/eddardd/MROT |
| Researcher Affiliation | Academia | Eduardo Fernandes Montesuma EMAIL Université Paris-Saclay, CEA, List, F-91120 Palaiseau France; Adel el Habazi* EMAIL École Centrale de Nantes, Nantes, France; Fred Ngolè Mboula EMAIL Université Paris-Saclay, CEA, List, F-91120 Palaiseau France |
| Pseudocode | Yes | Algorithm 1: Mass Repulsive Optimal Transport. |
| Open Source Code | Yes | Our code is publicly available at https: //github.com/eddardd/MROT |
| Open Datasets | Yes | We divide our experiments in 3 parts. Section 4.1 shows our results on Ad Bench (Han et al., 2022). Section 4.2 shows our experiments in fault detection on the Tennessee Eastman Process (Montesuma et al., 2024b; Reinartz et al., 2021). Furthermore, from Figure 9 (b), we see that our method struggles in high dimensional AD, such as those in the NLP datasets of Han et al. (2022). |
| Dataset Splits | Yes | In this experiment, we downsample the number of anomalous samples per fault category to {5, 10, , 30}. This results in a percentage of {4.45%, 8.53%, 12.28%, 15.73%, 18.92%, 21.87%} of anomalous samples. In Figure 11, we report our aggregated results over all percentage of anomalies. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used for running its experiments. It mentions computational complexity but no specific CPU/GPU models or configurations. |
| Software Dependencies | No | The paper mentions tools like POT (Python Optimal Transport), XGBoost, and algorithms like Simplex and Sinkhorn, but it does not specify version numbers for any software dependencies. |
| Experiment Setup | Yes | Here, we analyze the robustness of our method with respect to the entropic regularization penalty ϵ, and the number of nearest neighbors k in Nk. In our experiments, we evaluated our method on the values ϵ {0, 10 2, 10 1, 100}, where ϵ = 0 implies the use of exact OT, that is, linear programming. For MROT, we use k {5, 10, 20, , 50}. |