Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes
Adequacy of the bibliography: Yes

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

The paper presents a technically sound and well-organized study on semantic association rule mining in IoT, combining static knowledge graphs with dynamic sensor data. The proposed Autoencoder-based Neurosymbolic ARM method effectively reduces the number of rules while maintaining full data coverage. The experimental evaluation is thorough and provides strong support for the approach.

Strengths:
- Novel and well-justified integration of semantic and sensor data for rule mining.
- Comprehensive evaluation across diverse ARM methods and datasets.
- Clear and relevant discussion of scalability, execution time, and potential extensions.

Issues Addressed in the Revision:
- The manuscript now conforms to the journal’s formatting and layout requirements.
- The methodology and pipeline are more clearly explained, improving accessibility.
- The authors have added a discussion on model generalizability.
- The section on real-world scalability has been expanded.
- Practical implications, including integration into digital twins, are now explicitly addressed.
- The revision effectively resolves the initial concerns and enhances the clarity and applicability of the work.

Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

Sorry for the late review of the revision. I very much appreciate the inclusion of the running example. This makes the whole procedure much clearer. In particular, I noted that classes are used like categorical/nominal data ranges (exclusive values per feature), not like classes generally in OWL/RDFS ontologies (per default, these are not exclusive, but on the contrary, often part of a class hierarchy).

Particular thanks also for the highlighting of the changes and the detailed discussion of the reviews.

One remark just to make sure I understand correctly:
- In Equation 3 and 4, the "c_i" appear to be used to specify the count of different values in the range of a feature f_i, correct?

When looking through the bibliography, I found some items where the references are not up to journal standard. Examples are:

* Goldberg DE (2013) Genetic algorithms. pearson education India. --> Besides the typo, is it possible that this should rather refer to a textbook called "Genetic Algorithms in Search, Optimization, and Machine Learning" published by Addison Wesley?
* Gruber T (1993) What is an ontology. --> This reference indirectly "points" (via google scholar) to a web page that has already disappeared.
* Kennedy J and Eberhart R; and Khedr et al: "IEEE" was formatted as "Ieee" and "ieee"

Typographic remarks:

- "lost function" typo on p.5 is still present (should be "loss function")
- p. 4 To enhance readibility, better write "e(e_k) = (v_i, v_j) or e(e_k) = (v_j ,v_i) --> Otherwise the lone appearance of "(v_j ,v_i)" is confusing
- p. 6 "a Neurosymbolic approach" --> lowercase the "n"
- p. 8 "input to the trained Autoencoder represents consequent" -> Add an article before "consequent", .e.g. "the consequent"

Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

The author's response to the reviews were really helpful in understanding the paper. A few suggestions:

(1) The major hypothesis in the paper is that using an under-complete AE helped in extracting only the significant rules. It would be really nice if this hypothesis can be shown empirically by experimenting with an over-complete AE and showing that the rules extracted from under complete-AE is a sub-set of the rules extracted from over-complete AE.

(2) It would be nice to see the relationship between the rules extracted by the proposed method and the baselines. Are the rules extracted by the proposed method a subset of the rules extracted by the baseline? Or, are these rules not extracted by the baselines but is identified only by the proposed method? If so, could you present coverage or support of the rules exclusively extracted by the proposed method separately? A similar presentation is useful when comparing under-complete vs over-complete AE.