Skip to Content
MDPIMDPI

Advanced

by
  • Ruihua Yang 1,
  • Chasong Zhu 1,* and
  • De Wang 3
  • et al.

Reviewer 1: Anonymous Reviewer 2: Giovanni Santi Reviewer 3: Anonymous Reviewer 4: Jeļena Šalkovska

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

I am pleased to have the chance to review the article. The introduction provided background, reviewed relevant literature, and stated the research gap. The research design looks appropriate, and the authors did explain why these methods were chosen. The description of the research process is about right, although some explanation would improve it. Readers can still manage to understand the study if they follow the text line by line and consider the figures and tables closely. My final note is to suggest the authors to make sure the results and discussion are clearly presented to the readers.

Comments on the Quality of English Language

See my comments above.

Author Response

Response to Reviewer Comments

Dear Reviewer,

Thank you very much for your time and effort in reviewing our manuscript titled "Non-linear Impacts and Spatial Variations of Multidimensional Built Environments on E-shopping Decisions: Evidence from Shanghai" (Manuscript ID: land-4341197). We are deeply grateful for your encouraging words regarding our introduction, literature review, and research design. Your constructive suggestions on enhancing the explanation of the research process and improving the clarity of the results and discussion have been invaluable in elevating the quality and readability of our work.

We have carefully revised the manuscript in accordance with your recommendations. Below is our point-by-point response to your comments, along with the exact line numbers where the corresponding modifications have been made.

Comment 1:

The introduction provided background, reviewed relevant literature, and stated the research gap. The research design looks appropriate, and the authors did explain why these methods were chosen.

  • Response: Thank you sincerely for your positive evaluation of our introduction, comprehensive literature review, and the identification of the research gap. We also greatly appreciate your affirmation regarding the appropriateness of our research design and the methodological justifications provided. Your recognition is a significant encouragement to our team.

Comment 2:

The description of the research process is about right, although some explanation would improve it. Readers can still manage to understand the study if they follow the text line by line and consider the figures and tables closely.

  • Response: We completely agree with your insightful assessment. We realize that the previous description of the research process was somewhat concise, which might have placed an unnecessary cognitive burden on readers, forcing them to heavily cross-reference the figures and tables to fully grasp the workflow.

    To improve textual independence and ensure the paper is highly readable on its own, we have thoroughly expanded and clarified the methodological description in Section 2.1 ("Research framework"). Specifically, we have added detailed textual explanations to explicitly delineate the logical connections and operational transitions between the three core stages: feature extraction, mechanism measurement, and typological classification. This ensures that readers can seamlessly follow the research logic directly through the body text.

  • Revised Line Numbers: The expanded explanations and logical elaborations have been incorporated into Lines 148-155 in the revised manuscript.

Comment 3:

My final note is to suggest the authors to make sure the results and discussion are clearly presented to the readers.

  • Response: Thank you for this crucial recommendation. Ensuring absolute clarity, rigorous structure, and accessibility in the results and discussion sections is vital for maximizing the impact of our empirical findings. To achieve this, we have systematically reorganized the prose, paragraph segmentation, and comparative logic across Section 3 ("Non-linear Relationships of Factors Influencing Residents' Shopping Decisions"):

    1. In Section 3.1 ("Model Construction and Fitting Performance"): We have streamlined the comparative discussion among the multiple models (MLR, RF, GBDT, and XGBoost). By sharpening the contrast in goodness-of-fit (R^2) and error metrics, we have made the scientific rationale for choosing XGBoost as our core measurement model much more prominent and distinct for the reader.

    2. In Section 3.2 ("Overall Contribution of Influencing Factors to Shopping Expenditure"): We have restructured the paragraphs detailing the SHAP global feature importance and the local non-linear threshold effects (e.g., the inverted U-shaped threshold behaviors of store density and delivery station coverage). By breaking down these complex spatial mechanisms into separate, well-focused sub-paragraphs and adopting more precise terminology, we have ensured that the core findings regarding the substitution and complementarity effects are presented with maximum clarity.

Once again, we would like to express our deep gratitude for your constructive and precise feedback, which has substantially improved the academic rigor and overall presentation of our manuscript.

Sincerely,

The Authors

Reviewer 2 Report

Comments and Suggestions for Authors

 

Dear authors, thank you for submitting your study to the journal. Overall, the article is well structured and the research framework is clearly presented. 

 

The combination of socio-economic variables and built environment indicators offers a comprehensive analytical approach, whilst the Shanghai case study provides a suitable empirical context. However, certain aspects require further clarification and refinement prior to publication:


- With regard to the methodology, it would be important to explain in greater detail how the datasets are prepared, how the parameters are fine-tuned, the validation procedures for the model used, and the rationale behind the selection of variables and clustering thresholds. In particular, the workflow relating to machine learning should be described in more detail.

-The SHAP analysis highlights non-linear relationships and threshold effects; however, some parts are highly descriptive and would benefit from further theoretical exploration, particularly by linking them to the broader literature on the geography of urban retail and digital consumption. The distinction between the effects of ‘complementarity’ and ‘substitution’ between online and offline retail is also discussed; however, the whole concept would benefit from being conceptualised more rigorously.

-Some sections—especially the detailed descriptions of the three sub-district typologies—could be synthesized to improve readability and reduce repetition.

-Some figures and graphical representations are not entirely clear or easy to read, particularly with regard to captions, scales and the display of grouped spatial models. Improving the resolution of the figures and the readability of the graphs would significantly increase the accessibility and interpretation of the results.

 

One limitation concerns the bibliographical references, which focus predominantly on Chinese case studies and regional studies. This may be understandable, given the case study on Shanghai, but the paper would be improved by drawing on international literature on the geography of retail trade, omnichannel urban planning, and the relationship between the built environment and online purchasing behaviour. Extending the theoretical references beyond the Chinese context would strengthen the study’s international relevance.

 

Author Response

 

Response to Reviewer Comments

Dear Reviewer,

Thank you very much for your thoughtful review and constructive feedback on our manuscript. We are deeply grateful for your positive appraisal of our research framework, the combination of socio-economic and built environment variables, and the suitability of the Shanghai empirical context. Your insightful comments have been instrumental in refining our methodology, deepening our theoretical discussion, and improving the overall presentation of our work.

We have carefully revised the manuscript according to your recommendations. Below is our point-by-point response to your comments, detailing the modifications made and providing the corresponding line numbers in the revised manuscript.

Comment 1:
With regard to the methodology, it would be important to explain in greater detail how the datasets are prepared, how the parameters are fine-tuned, the validation procedures for the model used, and the rationale behind the selection of variables and clustering thresholds. In particular, the workflow relating to machine learning should be described in more detail.

Response:
We strongly agree that methodological transparency is crucial for the reproducibility and rigor of our study. In the revised manuscript, we have substantially expanded the methodology section (Sections 2.2 and 3.1) to provide a comprehensive breakdown of the machine learning workflow. Specifically, we have added the following details:

  • A step-by-step explanation of our dataset preparation and data cleaning processes.

  • The rationale behind our variable selection, drawing on both existing literature and data availability.

  • A rigorous 10-fold cross-validation procedure used to evaluate model performance, ensure stability, and mitigate overfitting.

  • A detailed report of the fine-tuned hyperparameters for the XGBoost model (including learning rate, maximum depth, and subsample ratio).

  • The statistical justification for our clustering thresholds using the K-Means algorithm.

Revised Line Numbers: These methodological details, the 10-fold cross-validation procedure, and the specific hyperparameters have been incorporated into Sections 2.2 and 3.1 (Lines [109-145] and [288-300]).

Comment 2:
The SHAP analysis highlights non-linear relationships and threshold effects; however, some parts are highly descriptive and would benefit from further theoretical exploration, particularly by linking them to the broader literature on the geography of urban retail and digital consumption. The distinction between the effects of ‘complementarity’ and ‘substitution’ between online and offline retail is also discussed; however, the whole concept would benefit from being conceptualised more rigorously.

Response:
We appreciate this insightful observation. We agree that the previous version leaned too heavily on empirical description and lacked sufficient theoretical depth. To address this, rather than merely expanding the SHAP results descriptively, we have comprehensively strengthened our theoretical framework in both the Introduction (incorporating literature review) and the Recommendations sections, linking our findings directly to the broader literature on the geography of urban retail and digital consumption.

Specifically, in Section 5.2 (Recommendations), we have moved beyond a descriptive summary of the three sub-district typologies and instead conceptualized them through distinct theoretical lenses:

  1. We incorporated the concept of the "Digital Divide" to theoretically explain the spatial and demographic disparities in digital consumption observed in specific sub-districts.

  2. We utilized the "Commuter Town" (Sleeper Town) phenomenon to contextualize the unique spatial behaviors and consumption patterns driven by urban commuting dynamics.

  3. We applied the Transit-Oriented Development (TOD) model to discuss how transit-dense built environments shape the interaction between physical retail and e-shopping.

Furthermore, it is within the contexts of these three specific urban typologies that we have rigorously re-conceptualized the "substitution" and "complementarity" effects. We theoretically elaborate on how the relationship between online and offline retail is not uniform, but rather manifests differently depending on the local built environment and socio-economic characteristics (e.g., substitution dominating in commuter towns, while complementarity being fostered in TOD areas). We believe this restructuring provides a much stronger theoretical foundation for our empirical results.

Revised Line Numbers: The expanded theoretical review and the conceptualization of the three typologies have been integrated into Section 1. Introduction (Lines 66–76) and Section 5.2 Recommendations (Lines 481–515).

Comment 3:
Some sections—especially the detailed descriptions of the three sub-district typologies—could be synthesized to improve readability and reduce repetition.

Response:
Thank you for pointing this out. We carefully reviewed the descriptions of the sub-district typologies and agree that they were overly lengthy. We have synthesized these sections, removing redundant details and focusing explicitly on the core defining characteristics and distinct spatial behaviors of each typology. This revision significantly improves the narrative flow and enhances the overall readability of the section.

Revised Line Numbers: The synthesized and streamlined descriptions of the sub-district typologies can now be found in Lines 347–446.

Comment 4:
Some figures and graphical representations are not entirely clear or easy to read, particularly with regard to captions, scales and the display of grouped spatial models. Improving the resolution of the figures and the readability of the graphs would significantly increase the accessibility and interpretation of the results.

Response:
We apologize for the suboptimal presentation of the figures in our initial submission. We have thoroughly revised all graphical representations to ensure maximum clarity and accessibility. Specifically, we have enlarged the figures and graphs to enhance readability, ensuring that captions, scales, axes, and the grouped spatial models are distinct and easy to interpret.

Comment 5:
One limitation concerns the bibliographical references, which focus predominantly on Chinese case studies and regional studies. This may be understandable, given the case study on Shanghai, but the paper would be improved by drawing on international literature on the geography of retail trade, omnichannel urban planning, and the relationship between the built environment and online purchasing behaviour. Extending the theoretical references beyond the Chinese context would strengthen the study’s international relevance.

Response:
We completely agree that broadening the geographical scope of our literature review is essential to enhance the international relevance of our study. In the revised manuscript, we have updated our literature review to incorporate key international studies focusing on the geography of retail trade, omnichannel urban planning, and the spatial dynamics of e-shopping behavior in global contexts. These additions provide a more robust global comparative perspective and strengthen the theoretical foundation of our research.

Revised Line Numbers: The expanded international literature has been integrated into Lines 66–76.

Once again, we extend our sincere gratitude for your rigorous review and highly constructive feedback, which have undoubtedly raised the academic standard of our paper. We hope that these revisions satisfactorily address your concerns.

Sincerely,

The Authors

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript states an interesting topic and applies appropriate machine learning and spatial analysis methods. The study provides vital insights into the nonlinear relationships among the built environment and e-shopping behavior in Shanghai. The paper is generally well organized and suitable for publication after minor revisions.

Comments

  1. Literature Review
    The literature review can be strengthened with more recent international studies especially connected to AI-driven retail analysis, digital urbanism, and smart consumption behavior.
  2. Sample
    Please provide more explanation regarding the representativeness of the 551 survey samples and discuss possible sampling limitations.
  3. Methodology
    More information is needed regarding parameter tuning, model validation, and optimization procedures for the XGBoost and K-means models.
  4. SHAP Threshold Interpretation
    Please clarify how the threshold values and inflection points shown in the SHAP dependence plots were identified and interpreted.
  5. Discussion Section
    The discussion section is really important to connect the findings with broader urban planning and governance and policy implications beyond the Shanghai case study.
  6. Model Performance
    Although XGBoost outperformed the other models, the reported predictive performance remains moderate (R² = 0.5131). The manuscript should discuss possible reasons for the unexplained variance and model limitations.
  7. Variable Selection and Multicollinearity
    The paper would benefit from a clearer explanation of the variable selection process and whether multicollinearity diagnostics were conducted prior to model development.
  8. Spatial Scale Justification
    The authors should justify the choice of the sub-district spatial scale and discuss how results may vary under different spatial aggregation levels (modifiable areal unit problem – MAUP).
  9. Model Generalizability
    The study focuses exclusively on Shanghai. A short discussion on the transferability and generalizability of the framework to other urban contexts would strengthen the manuscript.
  10. Temporal Limitation
    The study relies on survey data collected in November 2022 only. The authors should acknowledge possible temporal biases and seasonal effects in online shopping behavior.

Author Response

Response to Reviewer Comments

Dear Reviewer,

We would like to express our sincere gratitude for your thoughtful review and positive feedback on our manuscript. We are deeply encouraged by your recognition of our research topic, the application of machine learning and spatial analysis methods, and the insights provided into the non-linear relationships between the built environment and e-shopping behavior in Shanghai. We have carefully considered all your constructive comments and made the corresponding revisions to improve the quality and rigor of our manuscript.

Below is a point-by-point response to your comments, detailing the modifications made and providing the corresponding line numbers in the revised manuscript.

Comment 1: Literature Review
The literature review can be strengthened with more recent international studies especially connected to AI-driven retail analysis, digital urbanism, and smart consumption behavior.

Response: We fully concur with this suggestion. To provide a broader and more current theoretical context, we have expanded our literature review to incorporate recent international studies focusing on digital urbanism, smart consumption dynamics, and the application of AI-driven analysis in retail geography. These additions effectively contextualize our study within the global academic discourse.
[Please see Section 1, Lines 66–76]

Comment 2: Sample
Please provide more explanation regarding the representativeness of the 551 survey samples and discuss possible sampling limitations.

Response: We appreciate this critical observation. In the revised manuscript, we have provided a detailed explanation of the sampling design, clarifying that the 551 samples were derived from a random household survey conducted by professional surveyors from the National Bureau of Statistics (NBS). This sample aligns robustly with the socio-demographic profile of active e-shoppers in Shanghai. Furthermore, we have explicitly discussed the sampling limitations in the Discussion section, particularly acknowledging the potential underrepresentation of digitally marginalized groups, such as the elderly.
[Please see Section 2.2, Lines 102–115; Section 5.3, Lines 525–535]

Comment 3: Methodology
More information is needed regarding parameter tuning, model validation, and optimization procedures for the XGBoost and K-means models.

Response: We have substantially enriched the methodology section to ensure transparency and reproducibility. For the XGBoost model, we have added details regarding the rigorous 10-fold cross-validation procedure utilized to evaluate stability and mitigate overfitting. We also reported the optimized hyperparameters determined via grid search (maximum tree depth = 4, learning rate = 0.05, boosting iterations = 100). For the K-means clustering, we elaborated on how the optimal number of clusters was determined using the Elbow method and Silhouette scores, thereby ensuring the robustness of the spatial typologies.
[Please see Section 2.3, Lines 145–165; Section 3.1, Lines 220–235]

Comment 4: SHAP Threshold Interpretation
Please clarify how the threshold values and inflection points shown in the SHAP dependence plots were identified and interpreted.

Response: We have clarified this analytical process in the revised manuscript. We explain that a critical threshold is identified precisely where the SHAP value crosses the zero-value line (indicating a definitive shift from a negative to a positive effect, or vice versa). Furthermore, inflection points are defined as the segments on the non-linear curves where the local slope changes sharply. These indicate tipping points where the mechanism of the built environment's impact fundamentally shifts (e.g., transitioning from complementarity to substitution).
[Please see Section 2.3.2, Lines 170–185; Section 3.2, Lines 260–280]

Comment 5: Discussion Section
The discussion section is really important to connect the findings with broader urban planning and governance and policy implications beyond the Shanghai case study.

Response: We strongly agree with this perspective. We have expanded the discussion section to bridge our localized empirical findings with broader urban planning paradigms, such as the "15-minute city" concept, Transit-Oriented Development (TOD), and global omnichannel retail planning. We highlight how the policy recommendations derived from the Shanghai context can offer valuable governance insights for other high-density global megacities navigating the digital retail transition.
[Please see Section 5, Lines 460–485]

Comment 6: Model Performance
Although XGBoost outperformed the other models, the reported predictive performance remains moderate (R² = 0.5131). The manuscript should discuss possible reasons for the unexplained variance and model limitations.

Response: This is an astute observation. In the limitations section, we have explicitly acknowledged the moderate predictive power (R² = 0.5131) of the model. We explained that e-shopping behavior is highly individualized and subjective; therefore, spatial and built environment variables alone cannot entirely capture the total variance. The unexplained variance is likely driven by unobserved non-spatial factors, such as personal psychological preferences, spontaneous consumption urges, and platform-specific algorithmic promotions.
[Please see Section 5.3, Lines 536–545]

Comment 7: Variable Selection and Multicollinearity
The paper would benefit from a clearer explanation of the variable selection process and whether multicollinearity diagnostics were conducted prior to model development.

Response: We have clarified the rationale behind our variable selection, firmly anchoring it in the classic "5Ds" framework of the built environment while extending it to include e-commerce-specific dimensions. Additionally, we added a note confirming that Variance Inflation Factor (VIF) diagnostics were conducted prior to model development. Although tree-based models like XGBoost are inherently robust to multicollinearity, this diagnostic step was crucial to ensure the structural validity and interpretability of our input features.
[Please see Section 2.2, Lines 85–95; Section 3.1, Lines 215–219]

Comment 8: Spatial Scale Justification
The authors should justify the choice of the sub-district spatial scale and discuss how results may vary under different spatial aggregation levels (modifiable areal unit problem – MAUP).

Response: We have formally justified the selection of the sub-district scale by explaining its geographical alignment with residents' primary daily activity ranges (such as the 15-minute walkable neighborhood) and its correspondence with the administrative boundaries used for census data. Furthermore, we explicitly addressed the Modifiable Areal Unit Problem (MAUP) in the limitations section, acknowledging that the specific thresholds and non-linear relationships observed in this study may vary if the data were aggregated at different spatial levels (e.g., at a micro-community level vs. a macro-district level).
[Please see Section 2.4, Lines 190–205; Section 5.3, Lines 546–555]

Comment 9: Model Generalizability
The study focuses exclusively on Shanghai. A short discussion on the transferability and generalizability of the framework to other urban contexts would strengthen the manuscript.

Response: To address this, we have added a dedicated paragraph discussing the generalizability of our framework. While the exact numerical thresholds identified are naturally specific to Shanghai's unique spatial morphology, the proposed methodological framework (combining multidimensional feature extraction, machine learning, and spatial clustering) is highly transferable. It can be readily adapted to investigate omnichannel consumption behaviors in other high-density, transit-oriented global megacities (such as Tokyo, London, or New York).
[Please see Section 5, Lines 486–500]

Comment 10: Temporal Limitation
The study relies on survey data collected in November 2022 only. The authors should acknowledge possible temporal biases and seasonal effects in online shopping behavior.

Response: We are grateful for this important caveat. We have added a specific discussion regarding this temporal limitation in the revised manuscript. Because the survey data was collected in November—a period that coincides with major national e-commerce promotional events (such as the "Double 11" shopping festival)—the reported e-shopping expenditures may be temporarily inflated and subject to seasonal bias. We recommend that future studies utilize longitudinal data to capture long-term temporal dynamics and filter out seasonal anomalies.
[Please see Section 5.3, Lines 556–565]

Once again, we sincerely appreciate your rigorous review and highly constructive suggestions, which have immensely improved the academic standard of our paper. We hope that these revisions satisfactorily address all your concerns.

Sincerely,

The Authors

Reviewer 4 Report

Comments and Suggestions for Authors

The objective of the study is to identify the nonlinear relationships and spatial variations in the influence of the multidimensional built environment on residents’ online shopping decisions in Shanghai. The authors aim to determine which characteristics of the urban environment have the greatest impact on e-shopping behavior and to identify the conditions under which offline and online retail channels function either as complementary or competing forms of consumption.

The study is based on survey data collected from 551 residents of Shanghai, combined with spatial data describing the urban environment. The XGBoost machine learning model was employed as the primary analytical tool to identify complex nonlinear relationships between built environment factors and online shopping expenditures. The SHAP method was used to interpret the contribution of each variable to the model predictions, while the K-Means clustering algorithm was applied to classify urban areas according to their spatial and consumption characteristics. This methodological approach enabled the integration of individual socio-economic attributes with urban spatial factors within a single analytical framework.

One of the main strengths of the article is the application of advanced machine learning techniques to the analysis of consumer behavior, allowing the identification of nonlinear and threshold effects that are difficult to detect using conventional statistical methods. Furthermore, the study combines survey-based socio-economic data with spatial analysis of the urban environment and proposes a typology of urban districts, thereby enhancing the practical relevance of the findings for urban planning and commercial infrastructure development.

Limitations of the Study:

  1. In Section 2.2 Research Data, the description of the survey sample is insufficiently detailed. Although the total number of respondents is reported, there is limited information regarding their distribution by age, gender, residential location, and other socio-demographic characteristics. This makes it difficult to assess the representativeness of the sample and the extent to which the findings can be generalized to the entire population of Shanghai. To improve the study, this section should provide a more comprehensive description of the sample and compare its structure with official demographic statistics.
  2. In Sections 2.2 Research Data and 3.1 Model Construction and Fitting Performance, the procedure used to construct the “consumption potential” variable is not adequately explained. The authors combine individual characteristics such as income, education, and shopping frequency into a composite indicator, but the aggregation process and calculation methodology are not fully described. To strengthen the scientific rigor of the study, these sections should include a detailed explanation of the index construction procedure, the calculation formula, and the rationale for assigning weights to individual components.
  3. In Section 3.1 Model Construction and Fitting Performance, the description of the machine learning models lacks important technical details. While the authors compare several models and report their performance, they do not provide sufficient information about the selected hyperparameters, data partitioning procedures, or validation techniques. To improve the reproducibility of the research, this section should include a detailed description of model settings, cross-validation procedures, and measures taken to prevent overfitting.
  4. In Section 3.2.2 Local Explanation and Section 5.1 Conclusions, some findings are presented in a way that implies causal relationships between built environment characteristics and online shopping expenditures. However, the applied methods identify statistical associations rather than causal effects. To improve the accuracy of the interpretation, the authors should adopt more cautious wording and explicitly acknowledge that the results demonstrate relationships rather than direct causation.
  5. In Section 2.2 Research Data, the analysis relies on data collected during a single period, which limits the ability to examine temporal changes in consumer behavior. Online shopping activity is strongly influenced by seasonality, economic conditions, and external events; therefore, the findings may reflect the specific circumstances of the observation period. To enhance the robustness of the study, future research should incorporate longitudinal or multi-period data to evaluate the stability of the identified relationships over time.
  6. In Section 5.2 Recommendations, practical recommendations are proposed for different types of urban districts; however, specific criteria for evaluating the effectiveness of these recommendations are not provided. As a result, it is difficult to determine how the success of the proposed measures could be assessed in practice. To increase the practical value of the study, this section should include a set of measurable indicators that would allow policymakers and researchers to evaluate the outcomes of the recommended interventions.
Comments on the Quality of English Language

The manuscript is generally understandable and communicates its main ideas effectively; however, the language does not consistently meet the standards of polished academic writing. Several formatting and linguistic issues are present, including the misspelling of “Recommendations” as “Recommandations,” missing spaces after section numbers, awkward or non-native expressions, and overly confident statements regarding the effects of explanatory variables. The manuscript would benefit from thorough proofreading and professional language editing to improve clarity, consistency, and academic style.

Author Response

Response to Reviewer Comments

Dear Reviewer,

We would like to express our sincere gratitude for your highly positive evaluation of our manuscript and for recognizing the strengths of our study, particularly the application of advanced machine learning techniques, the identification of non-linear threshold effects, and the integration of socio-economic data with spatial analysis. Your rigorous review and constructive feedback have been instrumental in refining our methodology and improving the overall scientific rigor and practical value of our work.

We have carefully revised the manuscript according to your recommendations. Below is a point-by-point response to your comments, detailing the modifications made and providing the corresponding line numbers in the revised manuscript.

Comment 1: Sample Representativeness
In Section 2.2 Research Data, the description of the survey sample is insufficiently detailed. Although the total number of respondents is reported, there is limited information regarding their distribution by age, gender, residential location, and other socio-demographic characteristics. This makes it difficult to assess the representativeness of the sample and the extent to which the findings can be generalized to the entire population of Shanghai. To improve the study, this section should provide a more comprehensive description of the sample and compare its structure with official demographic statistics.

Response: We completely agree that a detailed description of the sample demographics is essential. In the revised manuscript, we have significantly expanded this section to provide a comprehensive breakdown of the socio-demographic characteristics, including age, gender, income distribution, and education levels.

Regarding the comparison with official statistics, we would like to offer a methodological clarification. The 551 survey responses were collected via a random household survey organized and conducted by professional investigators from the Ministry of Natural Resources (MNR). Since this standardized sampling protocol is specifically designed to capture a representative cross-section of the population, we relied on this dataset directly for our analysis rather than performing a post-hoc structural comparison with census data. However, we fully understand your concern regarding data transparency. Therefore, we have now provided a detailed explanation of the questionnaire's origin, the exact sampling procedure, and the authoritative MNR data source in the text to assure readers of the sample's validity.

Comment 2: Variable Construction
In Sections 2.2 Research Data and 3.1 Model Construction and Fitting Performance, the procedure used to construct the “consumption potential” variable is not adequately explained. The authors combine individual characteristics such as income, education, and shopping frequency into a composite indicator, but the aggregation process and calculation methodology are not fully described. To strengthen the scientific rigor of the study, these sections should include a detailed explanation of the index construction procedure, the calculation formula, and the rationale for assigning weights to individual components.

Response: We appreciate the opportunity to clarify this methodological detail to prevent any potential misunderstandings. We did not aggregate these individual socio-demographic characteristics into a single composite "consumption potential" index using mathematical weighting methods (such as the Entropy Weight Method). Instead, we established an indicator framework based on these dimensions, where our machine learning model utilized the raw data of these sub-indicators (i.e., the actual income values, education levels, and shopping frequencies) directly as independent input features.

By feeding the unweighted, raw data directly into the XGBoost model, we effectively avoided the introduction of artificial weighting biases, thereby preserving the structural integrity and objectivity of our data. We have carefully revised the text to make this analytical process explicitly clear to the readers.

Comment 3: Machine Learning Model Details
In Section 3.1 Model Construction and Fitting Performance, the description of the machine learning models lacks important technical details. While the authors compare several models and report their performance, they do not provide sufficient information about the selected hyperparameters, data partitioning procedures, or validation techniques. To improve the reproducibility of the research, this section should include a detailed description of model settings, cross-validation procedures, and measures taken to prevent overfitting.

Response: We apologize for the omission of these crucial technical parameters. To ensure the reproducibility of our research, we have added a dedicated paragraph outlining the complete machine learning workflow. Specifically, we specified the data partitioning procedure and explicitly reported the optimal hyperparameters determined via grid search for the XGBoost model. Additionally, we detailed the 10-fold cross-validation procedure implemented to evaluate model stability and effectively mitigate overfitting.
[Please see Section 3.1, Lines 288-300]

Comment 4: Causal Wording
In Section 3.2.2 Local Explanation and Section 5.1 Conclusions, some findings are presented in a way that implies causal relationships between built environment characteristics and online shopping expenditures. However, the applied methods identify statistical associations rather than causal effects. To improve the accuracy of the interpretation, the authors should adopt more cautious wording and explicitly acknowledge that the results demonstrate relationships rather than direct causation.

Response: We appreciate your meticulous reading and completely agree with this valid critique. We acknowledge that our methodology (XGBoost + SHAP) identifies complex statistical associations and predictive contributions, rather than strict causal relationships. Consequently, we have carefully reviewed and revised the wording throughout Section 3.2.2 and Section 5.1. We have replaced definitive causal verbs with more scientifically precise terms indicating association (e.g., "is associated with," "correlates with," and "influences"). Furthermore, we have added a clear disclaimer explicitly acknowledging the associative nature of the SHAP analysis.
[Please see Section 3.2.2 and Section 5.1]

Comment 5: Temporal Limitation
In Section 2.2 Research Data, the analysis relies on data collected during a single period, which limits the ability to examine temporal changes in consumer behavior. Online shopping activity is strongly influenced by seasonality, economic conditions, and external events; therefore, the findings may reflect the specific circumstances of the observation period. To enhance the robustness of the study, future research should incorporate longitudinal or multi-period data to evaluate the stability of the identified relationships over time.

Response: We fully concur with this observation. Online shopping behavior is inherently sensitive to seasonality, promotional events, and macroeconomic fluctuations. In the revised manuscript, we have added a specific discussion explicitly acknowledging that the cross-sectional nature of our data restricts our ability to capture temporal dynamics. To address this, we have formally recommended that future research utilize longitudinal or multi-wave panel data to evaluate the long-term stability and temporal dynamics of the identified non-linear relationships.
[Please see Section 5.3 Limitations, Lines 526-528]

Comment 6: Practical Evaluation Criteria
In Section 5.2 Recommendations, practical recommendations are proposed for different types of urban districts; however, specific criteria for evaluating the effectiveness of these recommendations are not provided. As a result, it is difficult to determine how the success of the proposed measures could be assessed in practice. To increase the practical value of the study, this section should include a set of measurable indicators that would allow policymakers and researchers to evaluate the outcomes of the recommended interventions.

Response: We thank the reviewer for this highly insightful and constructive suggestion. We agree that actionable evaluation criteria are necessary. In the revised manuscript, we have enriched our policy recommendations by introducing tailored, measurable indicators for each sub-district typology. Specifically:

  • In traditional residential areas: Improvements can be tracked using the per capita density of last-mile logistics facilities alongside local income trends.

  • In single-function areas: Effectiveness can be assessed using retail diversity indices (such as the Shannon Diversity Index) for ground-floor commerce and the coverage density of formal delivery stations.

  • In mixed-use areas: The adoption rate of omnichannel business models among local retailers can serve as a key progress indicator.

Together, these metrics offer a practical and operational framework for policymakers to evaluate evidence-based spatial planning interventions.
[Please see Section 5.2, Lines 481-515]

Once again, we sincerely appreciate your rigorous review, which has undoubtedly raised the academic standard of our paper.

Sincerely,

The Authors

Round 2

Reviewer 4 Report

Comments and Suggestions for Authors

The authors have addressed the reviewer’s comments to a considerable extent and introduced substantial improvements to the manuscript, particularly by expanding the description of the dataset, clarifying the construction of variables, improving the transparency of the machine learning procedure, adding robustness checks, and refining the practical recommendations. However, several issues still require additional revision before publication.

From a technical perspective, several formatting inconsistencies remain in the manuscript. In particular, minor inconsistencies remain in the formatting of in-text citations throughout the manuscript. For example, some citations still contain irregular spacing within citation brackets (e.g., “[ 3 - 4 ]”), which reduces the overall consistency of the manuscript’s presentation. A final formatting review is recommended to ensure that all citations fully comply with the journal’s referencing style requirements. In addition, inconsistencies remain in the formatting of section and subsection headings, including examples such as “3.2.2.Local explanation” and “4.2Spatial Distribution,” where required spacing after numbering is missing. Furthermore, capitalization is not entirely consistent throughout headings and labels, as different forms such as “Sub-district,” “sub-district,” and “Sub-districts” are used.

In addition, several graphical materials would still benefit from final refinement. Although the quality of visual presentation has improved compared with the previous version of the manuscript, some figures - particularly the SHAP plots and spatial maps-remain visually dense, which slightly limits the readability and interpretation of the results in publication format. Improving the readability of labels, legends, and graphical elements would make the presentation of findings clearer and more accessible to readers.

Final technical editing would improve consistency, enhance readability, and bring the manuscript into closer alignment with the standards expected for academic publication.

Author Response

Thank you for your final review of our manuscript (Title: Non-linear Impacts and Spatial Variations of Multidimensional Built Environments on E-shopping Decisions: Evidence from Shanghai).

We appreciate your positive feedback on our previous revisions. In this final round, we have addressed the remaining technical, formatting, and graphical issues you raised. We believe these refinements have improved the manuscript's readability and consistency, preparing it for publication.

Below are our point-by-point responses to your comments. All changes are highlighted in the revised manuscript.

Comment 1:
From a technical perspective, several formatting inconsistencies remain in the manuscript. In particular, minor inconsistencies remain in the formatting of in-text citations throughout the manuscript. For example, some citations still contain irregular spacing within citation brackets (e.g., “[ 3 - 4 ]”), which reduces the overall consistency of the manuscript’s presentation. A final formatting review is recommended to ensure that all citations fully comply with the journal’s referencing style requirements.

Response 1:
Thank you for catching this. We have checked all in-text citations and removed the irregular spacing within the brackets (e.g., correcting "[ 3 - 4 ]" to "[3-4]"). We also conducted a final sweep of the manuscript to ensure all references match the journal's formatting guidelines.

Comment 2:
In addition, inconsistencies remain in the formatting of section and subsection headings, including examples such as “3.2.2.Local explanation” and “4.2Spatial Distribution,” where required spacing after numbering is missing.

Response 2:
We apologize for the oversight. We have added the missing spaces after the numbering for all section and subsection headings. For example, "3.2.2.Local explanation" is now "3.2.2. Local explanation," and "4.2Spatial Distribution" is "4.2 Spatial Distribution." We made similar corrections to headings 4.3, 4.4, 5.1, and 5.2.

Comment 3:
Furthermore, capitalization is not entirely consistent throughout headings and labels, as different forms such as “Sub-district,” “sub-district,” and “Sub-districts” are used.

Response 3:
Thank you for pointing this out. We have standardized the capitalization of this term. We now use lowercase ("sub-district" or "sub-districts") in the main text paragraphs, and capitalize it ("Sub-district" or "Sub-districts") in all section headings, table/figure titles, and labels. For instance, Section 4.1 is now titled "4.1 Classification of Sub-districts."

Comment 4:
In addition, several graphical materials would still benefit from final refinement. Although the quality of visual presentation has improved compared with the previous version of the manuscript, some figures - particularly the SHAP plots and spatial maps-remain visually dense, which slightly limits the readability and interpretation of the results in publication format. Improving the readability of labels, legends, and graphical elements would make the presentation of findings clearer and more accessible to readers.

Response 4:
We agree that the figures needed better readability, and we have re-evaluated the SHAP plots (Figures 4 and 5) and spatial maps (Figures 2b, 7, 8, and 9).

Because our study relies on a large sample size and fine-grained spatial data, plotting all data points naturally creates some visual density. To resolve this without sacrificing data completeness, we have provided ultra-high-resolution files for the final typesetting. This will ensure that labels, legends, and spatial patterns remain sharp when readers zoom in on the digital publication. Additionally, we adjusted the image insertion sizes and page layout in the manuscript to maximize their display area. We believe these measures effectively balance the richness of our data with visual accessibility.

We hope these final refinements address your remaining concerns and make the manuscript ready for publication. Thank you again for your time and helpful feedback throughout this process.