Abstract
Current AI trip planners often struggle to respect user-defined constraints, such as specific starting and ending locations, without distorting the overall route. This paper introduces a diffusion-based guidance approach that "softly" steers the trajectory generation process. By iteratively injecting user queries during the denoising process of a Denoising Diffusion Probabilistic Model (DDPM), the proposed method ensures 100% faithfulness to user intentions while maintaining high-quality, geographically sensible recommendations.
Introduction
Trip recommendation systems aim to construct a sequence of Points of Interest (POIs) based on a user's query, such as a specific start and end point and a desired number of stops. While deep learning models like RNNs or GANs capture complex patterns, they often suffer from "unfaithfulness," where generated routes fail to meet the exact start/end constraints. Common post-hoc fixes, like "clipping-merging," often lead to trajectory distortion because the model does not account for the final destination during the generation process.
Methodology
The proposed approach leverages Denoising Diffusion Probabilistic Models (DDPMs) to iteratively refine a trajectory from noise. Instead of a single-pass generation, the model uses the iterative nature of diffusion to "steer" the output toward user constraints.
Key components include:
- Incremental Knowledge Injection: User constraints (start/end points) are injected during each denoising step.
- Soft Guidance: A weighting function $F(t) = e^{-\lambda t}$ adjusts the influence of user constraints. As the process approaches the final step ($t \to 0$), the weight of the user's specific points increases until they are perfectly integrated.
- Transformer Backbone: A Transformer layer learns the distribution of trajectory embeddings.
- Multi-task Learning: The model is trained with a joint loss function combining Mean Square Error (MSE) for diffusion and cross-entropy for POI classification.
Experimental Evaluation
The model was evaluated on two real-world datasets: Edinburgh and Glasgow. Performance was measured using $F_1$ scores for accuracy, and $S_{ratio}$ and $E_{ratio}$ to measure faithfulness to start and end points, respectively. Geographical sensible-ness was measured via $D_{med}$ (Manhattan distance error). The proposed method was compared against statistical baselines (Popularity, PoiRank) and deep learning models (DeepTrip, CTLTR, SelfTrip).
Results and Discussion
The proposed method significantly outperforms baselines in terms of faithfulness and trajectory quality.
As shown in the results table, the model achieves 100% $S_{ratio}$ and $E_{ratio}$ on both datasets, a feat not matched by existing deep learning models like SelfTrip or CTLTR. Furthermore, it achieves the lowest $D_{med}$ (0.411 for Edinburgh and 0.386 for Glasgow), indicating that the "soft" guidance effectively prevents path distortion. By injecting constraints gradually, the intermediate POIs "learn" to align themselves with the fixed endpoints, avoiding the awkward jumps seen in clipping-merging methods.
Conclusion and Future Work
By integrating user constraints directly into the diffusion generation process rather than as a post-processing step, this method ensures perfectly faithful trip recommendations. While the model shows superior accuracy and faithfulness, future work will address inference latency inherent in diffusion models and explore scalability to more complex constraints, such as specific time-of-day requirements.
Summary updated