Review Comment:
This paper proposed a framework that leverages defeasible reasoning with GeoKGs for geospatial similarity computation. The paper clearly written and generally well-organized. While the topic is timely and relevant, and I appreciate authors' efforts in addressing the challenges of geospatial similarity computation, I have significant concern regarding the clarity, rigor, and fairness of the experimental evaluation. Therefore, I do not recommend acceptance of this paper at its current stage.
1. Originality
The idea of introducing defeasible reasoning into geospatial similarity computation is conceptually interesting. However, the proposed approach relies primarily on predefined rules embedded in the GeoKG. While authors compare their method to KGE models and LFM-based models (e.g., Google Gemini), the comparison raises fairness concerns. In particular, LFM-based models are pretrained on vast external datasets, including geographic knowledge, while the proposed method relies on a limited internal rule base. This discrepancy in background knowledge is not sufficiently addressed, calling into question the comparability of the models. That's also one of the potential reasons why LFM-based models outperform. Furthermore, the paper does not clearly articulate what fundamentally new insights or technical contributions it introduces beyond existing rule-based KG systems.
The idea of introducing defeasible reasoning into geospatial similarity computation is conceptually interesting. However, the proposed approach relies primarily on predefined rules embedded in the GeoKG. While authors compare their method to KGE models and LFM-based models (e.g., Google Gemini), the comparison raises fairness concerns. In particular, LFM-based models are pretrained on vast external datasets, including geographic knowledge, while the proposed method relies on a limited internal rule base. This discrepancy in background knowledge is not sufficiently addressed, calling into question the comparability of the models. That's also one of the potential reasons why LFM-based models outperform. Furthermore, the paper does not clearly articulate what fundamentally new insights or technical contributions it introduces beyond existing rule-based KG systems.
2. Significance of the results
The reported results show that the proposed defeasible reasoning approach achieves lower accuracy than KGE models and the LFM-based models. While the authors argue that the interpretability compensates for lower accuracy, this trade-off is not compelling without a more substantial demonstration. There is no quantitative comparison of interpretability across models, which makes it difficult to assess whether the proposed method truly offers superior explainability. Tables 2 and 3 demonstrate the incremental performance gains through rule refinement and alternative conflict resolution strategies, but the overall accuracy remains lower than baseline models. Morever, the evaluation is limited to a single city (Amsterdam), and it's unclear how generalizable the system is to other regions.
There are also some minor things need to be addressed:
1.there should be full name of LFM when it's firstly mentioned.
2.Rule 3 introduces rating similarity as additional supporting evidence for semantic similarity. However, I find this connection somewhat unclear. Ratings primarly reflect customer satisfaction or preference, which may be influenced by many factors beyond cuisine - such as service or pricing. Therefore, it is not direct evidence how a small difference in rating meaningfully contributes to assessing whether two restaurants are semantically similar, especially in terms of their cuision type. In other words, likeness in customer ratings does not necessarily imply similarity in function or semantics, and this assumption should be better justified or empirically supported in the paper.
|
