TRACE: Traceroute-based Internet Route change Analysis with Ensemble Learning
Authors: Raul Suzuki, Rodrigo Moreira, Pedro Henrique A. Damaso de Melo, Larissa F. Rodrigues Moreira, Flávio de Oliveira Silva
Organizations: Institute of Exact and Technological Sciences – Federal University of Viçosa (UFV), Rio Paranaíba – MG – Brazil · Department of Informatics – School of Engineering, University of Minho (UMinho) – Braga – Portugal
Detecting Internet routing instability is a critical yet challenging task, particularly when relying solely on endpoint active measurements. This study introduces TRACE, a MachineLearning (ML)pipeline designed to identify route changes using only traceroute latency data, thereby ensuring independence from control plane information. We propose a robust feature engineering strategy that captures temporal dynamics using rolling statistics and aggregated context patterns. The architecture leverages a stacked ensemble of Gradient Boosted Decision Trees refined by a hyperparameter-optimized meta-learner. By strictly calibrating decision thresholds to address the inherent class imbalance of rare routing events, TRACE achieves a superior F1-score performance, significantly outperforming traditional baseline models and demonstrating strong effective ness in detecting routing changes on the Internet.
Figures & tables
Approach
Route-change Detection
Ensemble / Advanced ML
Temporal Feature Eng.
Imbalance Handling
Control-Plane Independent
Public Data Compatible
[ Islam et al. 2024 ]
[ Li et al. 2022 ]
[ Lin et al. 2025 ]
[ Al-Qudah et al. 2020 ]
[ Schmid et al. 2025 ]
[ Li et al. 2025 ]
Table 1: Comparison with Related Work.
Figure 1: Stacked ensemble architecture (Phases 1–3) and additional baseline models (Phase 4) trained directly on the engineered features for comparison.
Category
Count
Percentage
Description
Total training
19,965,159
100.00%
Full training set
Class 0 (stable)
19,595,589
98.15%
Majority class
Class 1 (changed)
369,570
1.85%
Minority class (target)
Total testing
8,556,497
–
Unlabeled split (30%)
Imbalance ratio
1 : 53.02
Table 2: Dataset statistics and class distribution.
Hyperparameter
Search Range
Step
Optimal Value
Number of Estimators
[100,5000]
100
200
Learning Rate
[0.001,0.05]
Cont.
0.0409
Maximum Tree Leaves
[10,80]
2
10
Maximum Tree Depth
[3,10]
1
3
Feature Fraction
[0.6,0.9]
Cont.
0.6687
Bagging Fraction
[0.6,0.9]
Cont.
0.7547
Table 3: Hyperparameter search space and TPE-selected optimal values for LightGBM.
Figure 2: Computational footprint of all evaluated models.
Figure 3: Relationship between computational cost and statistical benefit across models.
Figure 4: Predictive performance and latency trade-off across all models.
Model
F1 Score
Accuracy
CatBoost
0.855
0.865
DecisionTree
0.745
0.755
Dummy
0
0.5
KNN
0.635
0.655
LightGBM
0.835
0.845
RandomForest
0.825
0.835
Table 4: Overall performance comparison across all models.