Path Representation Learning (PRL) for road networks has attracted growing interest because of its relevance to routing, travel time prediction and other applications. Existing PRL methods mainly exploit co‑occurrence between road segments and paths to learn a fixed vector, overlooking the diversity of traveler behaviors and the correlations among road segments within a region.\ \ This work proposes to learn distributional representations that capture both varied traveler behaviors and regional dependencies. The proposed TraveL framework uses a Transformer to encode a path sequence together with its departure time, producing a distributional representation $\mathbf{z}$ that can be decoded into possible on‑path traveler behavior samples $\{\mathbf{s}_i\}$.\ \ To model regional correlation, a regional attention mechanism computes attention weights $\alpha_{ij}$ among road segments belonging to the same region and injects these weights into the path encoding, strengthening intra‑regional interactions.\ \ During training, the Kolmogorov‑Smirnov (K‑S) test is employed: the distance $D_{KS}=\sup_x|F_{\text{sample}}(x)-F_{\text{gt}}(x)|$ between the sampled behavior distribution $F_{\text{sample}}$ and the ground‑truth distribution $F_{\text{gt}}$ is incorporated into the loss, encouraging the model to generate realistic behavior distributions.\ \ Experiments on synthetic and real‑world road networks show that TraveL outperforms state‑of‑the‑art baselines, achieving a 14.7% improvement in mean K‑S distance for travel‑time distribution estimation, a 16.7% reduction in MAE for path similarity prediction, and a 3.97% reduction in MAE for destination prediction.\ \ Review: By integrating multi‑view information (path sequence, temporal context, regional attention) and distributional learning, TraveL addresses the information gaps of traditional PRL methods. The strong empirical gains across multiple tasks demonstrate its effectiveness and potential for broader traffic‑related applications.