<![CDATA[Transportation mode information automatically derived from Global Positioning System (GPS) trajectories can help operators and planners understand public, individual, and walking mobility patterns. This study compared Logistic Regression, Linear Support Vector Machine, Decision Tree, Random Forest, and Light Gradient Boosting Machine to classify walk, bike, bus, car/taxi, and train/subway. The Microsoft GeoLife dataset was processed from 5.3 million labeled GPS points into 172,105 segments with a maximum duration of 120 seconds from 55 users. Validation used Stratified Group 5-Fold Cross-Validation based on user_id. Sensitivity analysis was conducted on one main hyperparameter of each algorithm, while ablation testing was used to select feature groups. The results showed that 20 kinematic and motion-dynamics features produced the best performance. Random Forest achieved the highest macro F1 of 0.722±0.114, accuracy of 0.764±0.085, and macro AUROC of 0.903±0.058, followed by LightGBM with a macro F1 of 0.703±0.085. In out-of-fold aggregation, grouping bus and train/subway as public transport produced an F1 of 0.724 and a public-segment proportion difference of 3.14 percentage points compared with actual labels. The Friedman test showed an overall ranking difference, but pairwise comparisons after Holm correction were not significant. The results confirmed that tree-based ensembles are effective for nonlinear GPS patterns, although bus and car/taxi errors must be controlled before operational decision-making.]]>
Copyrights © 2026
