Disentangling the latent representation of traffic conflicts: A Multi-Task Ordinal Bayesian Variational AutoEncoder approach

High-resolution vehicle trajectories allow the risk of traffic-conflict to be assessed over short observation intervals. In this setting, the results may comprise several ordinal correlated surrogate safety measures rather than a single binary label. The joint use of multiple surrogate safety measures is established in the conflict analysis and is not claimed as a contribution. We instead examine whether TTC, DRAC, and PSD can be retained as three indicator-specific ordinal outcomes while their dependence and distinct uncertainty quantities are represented in one model. The resulting Multi-Task Ordinal Bayesian Variational AutoEncoder (MT-OBVAE) couples a shared probabilistic latent representation with three cumulative-probit decoders whose ordered thresholds are guaranteed by construction. Encoder dispersion, expected conditional entropy, and posterior predictive mutual information are used for representation-level ambiguity, task-level predictive ambiguity, and epistemic uncertainty, respectively. Relative to an empirical-prior reference, MT-OBVAE shows clear improvements in macro-F1, QWK, OrdMAE, AUROC, Brier score, and NLL, whereas the Accuracy gains are limited. Relative to seven learned baselines, MT-OBVAE achieves the highest mean AUROC across all three risk tasks, but the numerical margins are modest and several hard-label and calibration metrics are tied or favor a baseline, particularly for DRAC. In the challenging PSD task, MT-OBVAE achieves an AUROC of 0.764 and an NLL of 0.899. Beyond conventional prediction metrics, the learned representation supports three complementary analyzes: visualization of latent-state organization, representation-level ambiguity assessment, and task-specific latent sensitivity analysis. These model-native diagnostics examine quantities not supplied by feature-attribution methods alone, but they do not replace post hoc explainable-AI tools or establish physical phase transitions or causal risk-generation mechanisms. The evidence supports a contemporaneous and uncertainty-aware conflict-risk assessment. Representation-level ambiguity is elevated in intermediate-speed and severe-braking conditions, while TTC/PSD and DRAC exhibit distinct task-specific latent association patterns.