Extreme events in armed conflicts: a perspective

Frontiers in Physics, Journal Year: 2025, Volume and Issue: 13

Published: Feb. 20, 2025

Armed conflicts are characterized by changes in the number of fatalities that may span different orders magnitude. A change fatalities, crossing an order magnitude, be accompanied surprise. The paper suggests estimating probability extreme fatality (i.e., surprise) using a physics-informed approach. More specifically, aims to estimate unseen event where increase death toll crosses perspective is presented and illustrated with single case: postmortem analysis 7 October 2023 attack on Israel.

Language: Английский

---

Behind the mask: Random and selective masking in transformer models applied to specialized social science texts

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(2), P. e0318421 - e0318421

Published: Feb. 21, 2025

Transformer models such as BERT and RoBERTa are increasingly popular in the social sciences to generate data through supervised text classification. These can be further trained Masked Language Modeling (MLM) increase performance specialized applications. MLM uses a default masking rate of 15 percent, few works have investigated how different rates may affect performance. Importantly, there no systematic tests on whether selectively certain words improves classifier accuracy. In this article, we train set classify fake news around coronavirus pandemic using 15, 25, 40, 60 80 percent random selective masking. We find that 40 both selective, within-category but has little impact overall This finding important implications for scholars looking build classifiers, especially those where one specific category is more relevant their research.

Language: Английский

---

A novel hyperparameter tuning method for enhanced intrusion detection in network security

Turkish Journal of Engineering, Journal Year: 2025, Volume and Issue: 9(3), P. 519 - 534

Published: March 9, 2025

Intrusion Detection Systems (IDS) are essential for ensuring the security of enterprise networks and cloud-based systems, as they defend against sophisticated evolving cyberattacks. Machine learning (ML) techniques have emerged effective tools to enhance IDS performance, addressing limitations traditional methods. This study proposes a novel hyperparameter tuning method ML-based IDS, leveraging NSL-KDD dataset with extensive feature selection preprocessing address data imbalance redundancy. The method, integrating adaptive refinement stochastic perturbation, optimizes classifiers such Random Forest (RF), Gradient Boosting (GB), Extreme (XGB), achieving both higher detection accuracy (99.90% RF) improved computational efficiency. approach excels due its dynamic adjustment parameter ranges controlled randomness, converging faster than Grid Search by reducing iterations up 87.5%. experimental results demonstrate that tree-based models, particularly RF, outperform others their ability model complex, non-linear patterns, enhanced proposed method. Measured in terms convergence speed, CPU time, memory usage, this proves suitable deployment real-time, resource-constrained environments, offering scalable efficient solution network security.

Language: Английский

---

From Failure to Fusion: A Survey on Learning from Bad Machine Learning Models

Information Fusion, Journal Year: 2025, Volume and Issue: unknown, P. 103122 - 103122

Published: March 1, 2025

Language: Английский

---

Evidencing the Impact of Misinformed and Disinformed Beliefs on Individual and Group Behaviors

Psychological Inquiry, Journal Year: 2025, Volume and Issue: 36(1), P. 49 - 56

Published: Jan. 2, 2025

Language: Английский

---

Advancing Machine Learning in Industry 4.0: Benchmark Framework for Rare-event Prediction in Chemical Processes

Computers & Chemical Engineering, Journal Year: 2024, Volume and Issue: unknown, P. 108929 - 108929

Published: Nov. 1, 2024

Language: Английский

---

RI2AP: Robust and Interpretable 2D Anomaly Prediction in Assembly Pipelines

Published: April 30, 2024

Predicting anomalies in manufacturing assembly lines is crucial for reducing time and labor costs improving processes. For instance, rocket assembly, premature part failures can lead to significant financial losses inefficiencies. With the abundance of sensor data Industry 4.0 era, machine learning (ML) offers potential early anomaly detection. However, current ML methods prediction have limitations, with F1-measure scores only 50% 66% detection, respectively. This due challenges like rarity anomalous events, scarcity high-fidelity simulation (actual expensive), complex relationships between not easily captured by traditional approaches. Specifically, these relate two dimensions prediction: predicting when will occur understanding dependencies them. paper introduces a new method called Robust Interpretable 2D Anomaly Prediction (RI2AP) designed address both effectively. RI2AP demonstrated on simulation, showing up 30-point improvement F1 measure compared methods. highlights its enhance automated manufacturing. Additionally, includes novel interpretation mechanism inspired causal-influence framework, providing domain experts valuable insights into readings their impact predictions. Finally, model deployed real setting assembling parts. Results from this deployment demonstrate promise pipelines.

Language: Английский

---

ADVANCED MACHINE LEARNING TECHNIQUES FOR SEISMIC ANOMALY DETECTION IN INDONESIA: A COMPARATIVE STUDY OF LOF, ISOLATION FOREST, AND ONE-CLASS SVM

Jurnal Lebesgue Jurnal Ilmiah Pendidikan Matematika Matematika dan Statistika, Journal Year: 2024, Volume and Issue: 5(1), P. 49 - 61

Published: April 30, 2024

This study presents a comprehensive comparison of three machine learning algorithms for anomaly detection within seismic data, focusing on the unique geographical and geological context Indonesia, region prone to frequent events. Local Outlier Factor (LOF), Isolation Forest, One-Class SVM were assessed using meticulously curated dataset from Indonesian Meteorology, Climatology, Geophysical Agency, standardized ensure consistent feature scale. Our analysis encompassed both statistical metrics visualizations evaluate performance each algorithm. The emerged as most effective method, achieving highest silhouette score, indicative its superior cluster formation clear distinction between inliers outliers. Forest also demonstrated strong with favorable score Davies-Bouldin index, suggesting isolation capabilities. In contrast, LOF algorithm showed less precision, indicated by lower scores higher potential challenges in distinguishing normal anomalous patterns. Statistical validation Kruskal-Wallis H-test confirmed significant differences distributions algorithms, p-value 0.0. Visualizations through PCA t-SNE reinforced quantitative findings, displaying demarcation anomalies unlike LOF.The findings underscore importance selecting appropriate methods data analysis, highlighting robustness such applications. implications this research are profound risk management, providing insights that enhance accuracy reliability earthquake prediction systems, which is vital regions high activity Indonesia.

Language: Английский

---

Despite the perception that it is rare, by-catch of marine mammals can lead to population decline

ICES Journal of Marine Science, Journal Year: 2024, Volume and Issue: unknown

Published: Oct. 30, 2024

Abstract By-catch, the unintentional capture of non-target species in fishing gear, is often described as a rare event when it pertains to marine mammals. Yet, greatest threat many megafauna species, especially small cetaceans. How can both these statements be true simultaneously? The adjective “rare” itself rarely defined precise and quantitative way. Data collection features are crucial understand processes leading perception rarity samples on by-catch mammals other protected, endangered threatened (PETS). Statistical principles design monitoring schemes must upheld ensure sample representativeness scaling up impact estimates level whole fisheries. Random allocation observers or technologies (e.g. Remote Electronic Monitoring) strengthened test hypothesis that not only registered event, dedicated PETS by-catch. Even if at single operation, given large total number latter, their expansive spatial temporal extent, may still significantly mammal populations.

Language: Английский

---

A limit formula and recursive algorithm for multivariate Normal tail probability

Statistics and Computing, Journal Year: 2024, Volume and Issue: 35(1)

Published: Dec. 28, 2024

Language: Английский

---