Deep Learning (DL) shows great potential in personalised travel planning and traffic forecasting. However, existing techniques often produce inconsistent and unreliable results when handling complex travel constraints and traffic patterns, limiting their practical application in real-world scenarios. This thesis addresses some of these challenges by developing two complementary frameworks: TravLinkTo, which processes natural language travel queries through spatially-enhanced Large Language Model (LLM) to deliver reliable, constraint-satisfying travel itineraries for travellers and QuanTraffic, which quantifies uncertainty in transportation system traffic predictions by generating dynamic prediction intervals. These frameworks enhance the reliability and usability of DL applications in modern travel planning and traffic management.
First, TravLinkTo enhances LLM' spatial and temporal reasoning for automated travel planning. It implements a hierarchical framework that integrates LLM with external tools and knowledge bases, using knowledge graphs for intercity route optimisation and external data for intracity activity scheduling. Evaluations show that TravLinkTo outperforms alternative LLM + tooling approaches across base models and metrics, satisfying over 98% of spatial constraints, improving planning efficiency by 40%, and achieving higher user satisfaction with fewer hallucinations and constraint violations.
Next, QuanTraffic tackles uncertainty in traffic forecasting, such as speed and flow fluctuations during rush hours, which often lead to inaccurate arrival estimates and suboptimal planning. It introduces a model-agnostic uncertainty quantification framework that generates dynamically adapted prediction intervals for specific locations or traffic sensors, generating prediction intervals with reliable coverage probability for enhanced decision-making. Evaluations across five base models demonstrate that it outperforms six existing methods, achieving higher coverage with narrower intervals.
