 |
International Society of Science and Applied Technologies |
| Explainable AI for Road Safety: Identifying Hazardous Factors of Traffic Conflicts Using Neural Networks and SHAP Interaction Values | ||||
| Author |
Saki Yonezawa
|
|||
| Co-Author(s) |
Masashi Kuwano; Yuka Minamino; Mio Hosoe
|
|||
| Abstract | In this study, a predictive model is developed to
identify hazardous factors associated with traffic conflict risk in
the central area of Tottori City, Japan. Traffic conflicts are
operationalized using two complementary indicators: (i) traffic
accidents, representing observed collision events, and (ii) hardbraking
events, representing latent near-miss risks derived from
vehicle deceleration data. A neural network–based classification
model is employed to capture complex nonlinear relationships
between roadside environmental factors and traffic conflict
occurrence. Shapley Additive exPlanation (SHAP) interaction
values are applied to quantify both individual and interaction
effects among explanatory variables and to enhance
interpretability. Results indicate that network-related factors,
particularly betweenness centrality and its interaction effects,
play a dominant role in shaping traffic conflict risk. In addition,
demographic and built-environment factors such as population
distribution and building height influence traffic conflicts,
primarily via their interactions with network structure. These
findings demonstrate that traffic conflict risk arises from complex
interactions among multiple factors rather than from individual
variables in isolation. The proposed framework provides a datadriven
basis for identifying high-risk locations and supports the
design of targeted, proactive road safety interventions.
|
|||
| Keywords | Traffic conflicts, Hard braking events, Explainable AI, SHAP interaction | |||
| Article #: RQD2026-70 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |