|
Comparative modeling of women cyclists' perceived security using machine learning |
|---|---|
| รหัสดีโอไอ | |
| Title | Comparative modeling of women cyclists' perceived security using machine learning |
| Creator | Peyman Noorbakhsh |
| Contributor | Phromphat Thansirichaisree, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Machine learning, Ensemble learning, Deep learning, Deep neural network, Modeling travel behavior, Perceived security, Women cyclists, Bicycle, แมชชีนเลิร์นนิง, การเรียนรู้แบบ ensembles, ดีพเลิร์นนิง, โครงข่ายประสาทเทียมแบบลึก, การสร้างแบบจำลองพฤติกรรมการเดินทาง, การรับรู้ความปลอดภัย, นักปั่นจักรยานหญิง, จักรยาน |
| Abstract | Incorporating human behavioral factors into travel demand analysis represents an emerging trend in transportation planning, offering valuable insights for researchers in transportation engineering. Perceived security (PS), a key behavioral element, profoundly influences travelers' mode choices and demands, particularly for active modes like cycling and vulnerable users such as women in diverse cultural contexts, where its multifaceted nature heightens risks from social threats. This thesis addresses three interrelated research questions: (1) how women cyclists’ PS can be systematically predicted; (2) which socio-environmental factors most strongly influence PS; and (3) to what extent machine learning (ML) and deep learning (DL) can be adapted to small yet complex behavioral datasets. To answer these, a three-stage pipeline was developed on a modest dataset of 210 samples, collected via a VR-based bicycle simulator in Tehran, Iran, involving 55 female participants evaluating 16 socio-environmental scenarios. Employing 31 algorithms and techniques—comprising 24 ML algorithms, 4 ensemble types, and 3 advanced DL techniques—94 tuned models were developed, including AutoML tools (LazyPredict, TPOT, H2O), stacking ensembles, voting, stacking, bagging, boosting, single ML models (such as RF), and advanced DL architectures (such as ResNet). A custom stacking ensemble, DNN-stack2 (with LightGBM, KN, ET base learners and DNN meta-learner), achieved the highest performance at 84.07% accuracy and 85.32% macro F1-score. Notably, the Feature Tokenizer Transformer (FTT) yielded reliable results (79.2% accuracy, 78.6% macro F1) despite the dataset’s limited size, which typically challenges data-intensive DL models, demonstrating its efficacy in capturing complex interactions with low cross validation variance (71.5 ± 2.7%). A notable finding is the consistent feature importance across paradigms: crowdedness and lighting emerged as dominant predictors not only in DNN-stack2 and FTT but also in the RF baseline. This alignment serves as a proxy for the validity of both advanced and classical ML models, reinforcing confidence in the robustness of findings. The convergence of ML and DL results—despite their different methodological foundations—further highlights the potential of combining approaches. Such consistency also encourages the collection of larger datasets, where both DNN stack2 (ML-focused) and FTT (DL-focused) can be refined to deliver deeper insights for urban planners. This work contributes a reproducible pipeline for PS modeling, validates ML/DL approaches for small datasets, identifies key socio-environmental predictors of women cyclists’ security perceptions, and proposes practical applications such as interactive security dashboards. While the constrained sample size underscores the need for expanded, multi-regional data to enhance generalization and refine medium PS predictions, these findings advance gender-sensitive sustainable mobility research and provide actionable knowledge for planning safer cycling environments. |
