SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation

Thank you for your thoughtful and encouraging review. We sincerely appreciate your recognition of the novelty, technical soundness, and practical relevance of our work, as well as your positive comments on the manuscript’s clarity, experimental rigor, and reproducibility. We also thank you for your valuable time and effort in reviewing our paper. Below, we carefully address the weaknesses and questions you raised.

Response to Weaknesses: W1: Thank you for your suggestion. We have added a replacement-based ablation study to clarify the activation function choices. Specifically, we denote the activations in Eqs. (1), (2), and (10) as $\phi_{(1)}$, $\phi_{(2)}$, and $\phi_{(10)}$, and conduct exhaustive experiments by replacing each with LeakyReLU, SiLU, or GELU. The results are summarized in the three tables provided. Although some alternative combinations slightly outperform our original settings on individual metrics, our chosen configuration still achieves consistently strong performance across both datasets and metrics. Therefore, we retain the original setup.

$\phi\_{(1)}$ == LeakyReLU:

$\phi\_{(1)}$ == SiLU:

$\phi\_{(1)}$ == GELU:

We will include this new set of results and corresponding discussion in the revised appendix.

W2: Thank you for pointing this out. We will revise the mentioned lines and carefully proofread the entire manuscript to eliminate redundant statements and improve clarity.

W3: Thank you for the helpful suggestions. We acknowledge that the suggested recent works exhibit partial similarities with components of our framework. We will include these works [1–3] in the revised version and elaborate on their connections and distinctions with our method in the related work section.

Response to Questions:

Q1: Thank you for the valuable suggestion. The Yelp dataset covers data up to the year 2021, and the timestamps in the dataset (e.g., 1559768993) are recorded as Unix epoch time in seconds, which can be converted to standard date-time format for preprocessing. To improve clarity for readers, we will include these detailed descriptions in the revised manuscript.

Q2: Thank you for the insightful suggestion. Extending the SPOT-Trip framework to generate top-K candidate trajectories would indeed require incorporating appropriate search and trajectory scoring strategies to balance diversity and relevance. We consider this an important direction for future work and plan to explore it in subsequent research.

Thank you for your constructive and detailed feedback on our paper. We sincerely appreciate your time and effort in reviewing our manuscript. We are grateful for your recognition of the motivation, methodology, and case study design. Below, we will address the weaknesses and questions you raised.

Response to Weaknesses:

W1: Thanks for pointing this out. We apologize for the oversight. In the revised version, we have carefully proofread the manuscript and added clear definitions for all symbols, including d (i.e., the dimension of the hidden representation), and corrected the punctuation in all equations, including Eq. (3) and Eq. (8).

W2: Thanks for the insightful suggestions. In the revised version, we will streamline Sections 1 and 3 to remove redundant content and improve clarity. Additionally, we agree that the Related Work section is important and will move it from the appendix to the main text. The statistics of the knowledge graph are already provided in the appendix--we will also relocate them to the main text for better visibility.

W3: Thanks for pointing this out. As you suggested, the proposed Static Aggregator indeed functions as a mean pooling mechanism. To avoid introducing an unnecessary new concept, we will rename it as Mean Aggregator in the revised version. Regarding the baseline comparison, many existing out-of-town recommendation models [1][2][3] focus on next POI prediction rather than full trip generation, and thus are not directly applicable to our task. KDDC [3] is a relatively suitable and SOTA approach among these methods, so we have adopted its knowledge-enhanced module in our Base framework (Base+KDDC) as a comparative baseline.

W4: We apologize for the confusion. The paper [4] does not explicitly name its method, so we referred to it as PPROC based on the name used in its official GitHub repository. We will clarify this naming choice in the baseline descriptions to avoid misunderstanding.

W5: In fact, removing the Static-Dynamic Preference Fusion module is equivalent to discarding both the Knowledge-Enhanced Static Preference Learning and the ODE-Based Dynamic Preference Learning components (i.e., the Base baseline reported in our paper). To address your concern, we will explicitly clarify this equivalence in the revised version.

Response to Questions:

Q2: Yes, your understanding is correct. Existing trip recommendation methods [4][5] generally do not consider user check-in behaviors in their hometown, while existing out-of-town recommendation works [1][2][3] are mostly limited to next POI prediction in an unfamiliar city. To the best of our knowledge, no prior work has addressed the generation of a full trip in an out-of-town scenario. This is the gap our work aims to fill.

Q3: Thanks for the question. Our motivation for separating hometown and out-of-town check-in sequences follows prior out-of-town recommendation works [1][2][3], which emphasizes the interest drift between users' hometown preferences and their behaviors in unfamiliar cities. The goal of out-of-town recommendation is to suggest POIs that users might be interested in when traveling in out-of-town regions. Moreover, the regions in our study are defined exactly as annotated in the original datasets (Foursquare and Yelp), without any modification. For example, users whose hometown is Los Angeles and who travel to San Diego are considered to be on out-of-town travel. The case reflects actual travel behaviors rather than routine suburban commutes. Therefore, daily commuting--such as traveling from a city to nearby suburbs--is unlikely to be captured as out-of-town behavior, given the dataset's regional granularity. For more details on how we distinguish the hometown and out-of-town regions, please refer to our response to Q6.

Q4: Thanks for the question. In our task, the time intervals between check-ins during a user’s out-of-town trip are irregular and unknown. Therefore, we cannot directly use actual time gaps as input. RNNs and Transformers generally assume that input sequences are equally spaced or have known time gaps. When we use step indices corresponding to the steps to be predicted (e.g., POI 1 → POI 2 → POI 3), these models may implicitly assume a constant rate of preference change, which does not reflect the true dynamics of user behavior during travel. In contrast, our use of Neural ODEs allows the model to learn nonlinear, continuous preference transitions over a latent time dimension. This enables more flexible modeling of preference drift at variable speeds between steps. Empirically, this design also leads to better performance (see below or our response to reviewer 8nVL Q3), confirming the advantage of Neural ODEs in handling such irregular, implicit temporal dynamics.

Q5: Thank you for the question. The value may appear low at first glance. However, our task is fundamentally more challenging than next-POI prediction. Rather than predicting just the next check-in point, our model must generate an entire out-of-town trip and match multiple POIs in both content and exact position. This position-wise alignment across a sequence significantly increases difficulty, especially under real-world data sparsity. Despite the difficulty, compared to prior trip recommendation approaches, our method achieves a significant performance improvement under the same evaluation metrics. Regarding ACC, NDCG, and MRR using in traditional next POI prediction: These metrics usually rely on negative sampling, which are not applied in trip recommendation evaluation. More importantly, they are ranking-based and focus on whether the correct POI appears near the top of a list--not on position-wise sequence matching, which is essential in our trip recommendation task. Our goal is to assess whether the predicted sequence aligns step-by-step with the ground-truth travel trajectory. Therefore, following prior trip recommendation work [5][6], we use $F_1$ and $PairsF_1$, which better reflect sequential prediction quality.

Q6&Concern about user privacy in Limitations: Thanks for the question. As mentioned in the dataset section of our paper, we follow the hometown discovery approach introduced in prior work [2] to distinguish between a user's hometown and out-of-town regions. In brief, it first extracts a user's check-in sequences across different regions over continuous time windows. Then, based on factors such as the number of check-ins and time span, it assigns a score to regions and selects the highest-scoring one as the user's hometown. Detailed examples can be found in the dataset of our released code. As for location categories, we do make use of semantic POI types from the dataset (e.g., Restaurant, Mall), but the ‘home’ does not appear in the available taxonomy and is not used for identifying hometowns. Regarding the concern about user privacy, we would like to clarify that, consistent with prior out-of-town recommendation works [1][2][3], all user data used in this study has been fully anonymized. No personally identifiable information is involved, and the hometown identification is performed at the region level based solely on anonymous user IDs and aggregated check-in behavior.

Q7: As noted in our response to Q3, the motivation for treating out-of-town recommendation as a separate task lies in the drift between users’ preferences in their hometown and in unfamiliar regions. Traditional next-POI methods [7][8] typically model a user's check-ins within a single region (e.g., New York) and recommend the next POI in that same region. However, when users travel to a new region (e.g., from Los Angeles to San Diego), these models often fail to generalize due to region-specific patterns. While such methods may work for hometown prediction, they are insufficient for out-of-town scenarios, where behavior is driven more by travel preferences. Hence, distinguishing between hometown and out-of-town regions is necessary to capture this interest shift, as emphasized in prior works [1][2][3].

Q8: Thank you for the suggestion. We agree that user behaviors vary across regions, and visualizing underperformance is important. In the appendix (Case Study), we show that SPOT-Trip may generate repetitive POIs in some trips, reflecting a limitation in diversity modeling. While we do not conduct a systematic regional error analysis in this work, we see this as a valuable direction for future research.

Q9: Our work focuses on modeling user preferences in trip recommendation. Incorporating feasibility constraints (e.g., time, cost) is important for real-world use but typically requires extra data sources. We consider this an important direction and believe that combining preference learning with constraint-aware modeling is a promising future path.

Reference:
[1] Xin, Haoran, et al. "Out-of-town recommendation with travel intention modeling." AAAI. 2021.
[2] Xin, Haoran, et al. "Captor: A crowd-aware pre-travel recommender system for out-of-town users." SIGIR. 2022.
[3] Liu, Yinghui, et al. "KDDC: Knowledge-driven disentangled causal metric learning for pre-travel out-of-town recommendation." IJCAI. 2024.
[4] Iakovlev, Valerii, and Harri Lähdesmäki. "Learning Spatiotemporal Dynamical Systems from Point Process Observations." ICLR, 2025.
[5] Zhang, Jiale, et al. "Encoder-decoder based route generation model for flexible travel recommendation." TSC (2024).
[6] Shu, Wenzheng, et al. "Analyzing and mitigating repetitions in trip recommendation." SIGIR. 2024.
[7] Qin, Yifang, et al. "A diffusion model for POI recommendation." TOIS (2023).
[8] Zhou, Hailun, et al. "Disentangled Graph Debiasing for Next POI Recommendation." SIGIR. 2025.

We sincerely thank the reviewers for their timely responses. We would like to take this opportunity to clarify a few remaining concerns raised as follows.

More concerns about Q3: Thank you for the feedback. In the current study, we adopt the established distinction between hometown and out-of-town regions primarily to maintain consistency and comparability with prior research, which also facilitates fair evaluation. Our main contribution lies in extending the out-of-town recommendation task to the more challenging out-of-town trip recommendation task, and proposing a novel method to address it. Exploring more fine-grained user behavior segmentation methods--such as leveraging multimodal large models--would be highly beneficial for advancing this field, and we plan to consider such directions in our future work.

More concerns about Q4: Thank you for your comment. We believe some clarification is needed, as our original explanation may have been too brief due to space limitations.
Our task consists of two stages: model training and recommendation (inference). During training, the time intervals between check-ins in both regions are known and can be used as input. However, during recommendation, only the user’s hometown history is accessible. The timestamps of future out-of-town check-ins are unknown, making it impossible to use real time gaps as input. Consequently, the model must rely on discrete step indices (i.e., equal intervals) when generating out-of-town recommendations.
This leads to a discrepancy: RNNs or Transformers may leverage actual time gaps during training, but must operate with uniform out-of-town step indices at inference time. Such inconsistency can introduce temporal mismatch and potentially degrade performance, as these models are sensitive to the encoding of time.
In contrast, Neural ODEs model user preference transitions as continuous functions over latent time, which enables smoother generalization even when explicit intervals are unavailable during recommendation. This inductive bias toward continuous dynamics proves empirically beneficial in our task, as demonstrated in our ablation studies.

More concerns about Q5: Thank you for the comment. We would like to further clarify some misunderstandings. Acc@k, NDCG, and MRR indeed do not inherently rely on negative sampling. What we intended to convey is that some prior works often employ negative sampling when dealing with large candidate sets, in order to compute these metric values. Nevertheless, this was also not our main reason for not choosing these metrics. Our core point is that these ranking-based metrics are not well-suited for evaluating sequence-level prediction, which is the focus of our task. Specifically, our model, like the baselines we compare against, generates entire POI sequences through sampling-based methods (e.g., top-p (nucleus) sampling), rather than producing explicit top-k ranked lists. Therefore, sequence-level metrics such as F1 and PairsF1 provide a more appropriate and fair evaluation of the generated trajectories in this study.

We again sincerely thank your for carefully reading our article and providing us with useful comments and helpful criticism. We will be glad to address any further concerns if you arise.

Thank you for your detailed comments and feedback. We sincerely appreciate your time and effort in reviewing our manuscript. We will address the weaknesses and questions you raised below.

Response to Weaknesses and Questions:

W2&Q1: We agree that prior optimization-based approaches leveraging POI categories provide a way to generalize user preferences across regions. In fact, our method also makes use of POI category information, but in a more structured and expressive way. Specifically, as described in line 553 of our Appendix, we incorporate POI categories in the construction of a knowledge graph, where we generate entity-specific relations using multiple types of auxiliary information--including categories, star ratings, user reviews, and associated regions. This allows us to go beyond simple category matching and instead learn richer semantic representations of POIs through knowledge-enhanced embedding.
Moreover, relying solely on POI categories, as done in previous work, is insufficient to capture the nuanced interest drift [1] that occurs when users travel across regions--i.e., the same user may exhibit different check-in behaviors in different cities. To address this, we introduce a knowledge-enhanced preference alignment module that explicitly learns and adapts to such cross-region preference shifts. We hope this clarifies the distinction between our work and earlier methods, and highlights the novelty of our approach in addressing the unique challenges of out-of-town trip recommendation. We will include illustrative examples in the revised main text to clarify that POI categories have been incorporated into our framework.

W1&Q2: With respect to the PersTour method [2] mentioned by the reviewer, we acknowledge its relevance. However, due to its relatively early publication and the fact that it has been consistently outperformed by more recent methods such as GraphTrip [4], MatTrip [5], and AR-Trip [6] (which we already included as baselines), we initially did not include it in our comparison. Nevertheless, in response to the reviewer's suggestion, we have now additionally implemented and compared PersTour with our approach. The results are presented in the updated table below. The results further confirm the effectiveness of our proposed method over earlier approaches. We will also update Table 1 in the revised manuscript to include these new comparison results.
As for the MDTRP method [3] mentioned by the reviewer, we note that the authors have not released the source code, either in the paper or through public platforms such as GitHub or Kaggle, which makes it difficult to reproduce their approach within a short timeframe. However, we observed that the overall architecture of MDTRP is highly similar to that of MatTrip [5]--a baseline we have already included in our experiments. Both models adopt LSTM-based encoders and attention mechanisms, and both learn user interests over POI categories. As shown in Table 1 of our paper, our method consistently outperforms MatTrip, which indirectly demonstrates the advantage of our approach over methods in the style of MDTRP.

Q3: Thank you for this thoughtful question. In our current task formulation, we focus on proactively generating personalized trip trajectories based on a user's historical hometown records and their issued query, which specifies a starting point, an ending point, and the desired number of intermediate POIs. While we acknowledge the importance of considering travel time and distance, not all users provide explicit budget constraints for these factors in practice. As such, our current framework does not assume such inputs, and instead recommends POI sequences based primarily on user preference modeling. Nevertheless, we agree that evaluating and optimizing tour quality in terms of diversity, total travel distance, and estimated time is a valuable future direction. In future work, we plan to incorporate additional modules to account for user-specified other constraints (e.g., time or distance budgets) and to design corresponding evaluation metrics that better capture the practical quality of generated trips.

A minor typo: Thank you for pointing out this typo. We will correct it in the revised version.

Reference:
[1] Xin, Haoran, et al. "Out-of-town recommendation with travel intention modeling." AAAI. 2021.
[2] Lim, Kwan Hui, et al. "Personalized tour recommendation based on user interests and points of interest visit durations." IJCAI. 2015.
[3] Huang, Feiran, Jie Xu, and Jian Weng. "Multi-task travel route planning with a flexible deep learning framework." TITS (2020).
[4] Gao, Qiang, et al. "Dual-grained human mobility learning for location-aware trip recommendation with spatial-temporal graph knowledge fusion." Information Fusion (2023).
[5] Zhang, Jiale, et al. "Encoder-decoder based route generation model for flexible travel recommendation." TSC (2024).
[6] Shu, Wenzheng, et al. "Analyzing and mitigating repetitions in trip recommendation." SIGIR. 2024.

Thank you for the positive and encouraging evaluation of our work. We are especially grateful for your recognition of the practical value of the problem, the technical soundness of our framework, the thoroughness of our experiments, and the quality of both the writing and codebase. We also sincerely appreciate your constructive feedback and the time and effort you dedicated to reviewing our manuscript. Below, we carefully address the weaknesses and questions you raised.

Response to Weaknesses:

W1: While both our work and KDDC [1] address out-of-town scenarios, we would like to emphasize that our task formulation is fundamentally different from that of KDDC in terms of input conditions and recommendation granularity. Specifically, KDDC addresses a pre-travel decision-making problem, where the system recommends the next out-of-town POI from unknown regions to users who have not yet selected a travel destination. The goal is to suggest which POI to visit next, given the user's historical hometown records. In contrast, our work targets a post-destination trip recommendation task. We aim to recommend a sequence of POIs (i.e., a full trip trajectory) to users who have already chosen (or arrived at) a specific destination region. Our goal is to help these users plan meaningful intra-region journeys in response to user-issued queries that specify a starting point, an ending point, and a desired number of POIs to visit. As illustrated in Fig.1 of our paper, our setting reflects a significantly different application scenario that prior out-of-town recommendation works [2][3], including KDDC, do not cover, and the two problems serve users with distinct needs at different stages of their travel decision pipeline. We hope this clarification addresses the concern regarding the novelty of our problem definition.

W2: We acknowledge that a small portion of our method illustration (the design and color scheme of the semantic knowledge aggregation module) bears some resemblance to KDDC. This is because such an aggregation module is a common and useful component in knowledge-driven recommender systems [4][5]. Beyond this overlap, our methodology in technical implementation differs from KDDC's. While KDDC employs a segmented pre-training scheme and a knowledge aggregation module to enhance POI representations before applying disentangled causal learning to recommend the next out-of-town POI, our approach takes a different route. We adopt a more lightweight strategy by using TransE-based alternate training to integrate knowledge, and we restrict this component solely to modeling users’ static preferences.
More importantly, our framework introduces a novel ODE-based module to capture dynamic user preference transitions in out-of-town regions—another critical capability for our sequential trip recommendation setting. Compared to KDDC’s task, which only requires identifying a single POI that matches the user's static out-of-town preference, our task involves modeling how user preferences evolve over time and location within a destination region, making dynamic modeling essential. To better illustrate the difference, a tabular is included in our response to Q1.
Moreover, in Tables 1, 2, and 4 of our paper, we include several variants where we replace our knowledge-enhanced module with the one from KDDC. The consistent improvements across both datasets show that our approach is more effective.

Response to Questions:

Q1: To supplement our responses to W1 and W2 above, we summarize the distinctions below:

To further address your concern, we will incorporate explicit discussions of these differences in the Related Work and Baseline sections of the revised manuscript.

Q2: In addition to the above validation from the out-of-town recommendation perspective that we are the first to study the out-of-town trip recommendation problem, we further justify our novelty from the trip recommendation perspective. Prior trip recommendation methods have not yet distinguished local users from those arriving from other regions, nor solved the preference drift that occurs when a user travels outside their hometown. Therefore, we introduce the out-of-town trip recommendation task and devise a dedicated method to address it.

Q3: Thanks for the question. We choose Neural ODEs not for novelty, but because they offer continuous-time modeling capabilities that are critical in our task. Unlike RNNs and Transformers that operate over discrete and often fixed time steps, Neural ODEs, when integrated with point process tools, allow us to model travel trajectories with irregular time sampling from a probabilistic perspective, enabling a flexible characterization of continuous-time user out-of-town dynamic preferences.
Following the reviewer’s suggestion, we have implemented two additional temporal modeling variants: Base+KSPL+GRU (RNN-based) and Base+KSPL+Transformer. In both variants, we retain the overall architecture of our model, replacing only the ODE-based dynamic preference learning component with GRU and Transformer decoders, respectively. Specifically, both variants autoregressively generate the user’s dynamic preferences over a fixed number of time steps. As shown in the following table, our ODE-based method (SPOT-Trip) consistently outperforms both the RNN variant (Base+KSPL+GRU) and the Transformer variant (Base+KSPL+Transformer) on both datasets across all evaluation metrics:

This demonstrates the superiority of Neural ODEs in modeling continuous-time dynamics, which is particularly suitable for sparse and irregularly sampled POI trajectory in out-of-town scenarios. We will include these additional results in Table 4 of our revised manuscript to substantiate our design choices. Thank you once again for this valuable and constructive suggestion.

Reference:
[1] Liu, Yinghui, et al. "KDDC: Knowledge-driven disentangled causal metric learning for pre-travel out-of-town recommendation." IJCAI. 2024.
[2] Xin, Haoran, et al. "Captor: A crowd-aware pre-travel recommender system for out-of-town users." SIGIR. 2022.
[3] Xin, Haoran, et al. "Out-of-town recommendation with travel intention modeling." AAAI. 2021.
[4] Yang, Yuhao, et al. "Knowledge graph self-supervised rationalization for recommendation." KDD. 2023.
[5] Yang, Yuhao, et al. "Knowledge graph contrastive learning for recommendation." SIGIR. 2022.