Review Reports - Link Prediction Using Temporal Graph Neural Network Model

Round 1

Reviewer 1 Report (Previous Reviewer 3)

Comments and Suggestions for Authors

This paper applies TGNN to propose a method for analyzing social networks. Although the experimental results appear to be satisfactory, the claim of methodological novelty is insufficient. There are also several parts where the academic format is not properly maintained. Therefore, substantial revisions are required in order for readers to fully understand the usefulness of the authors’ approach.

Major Comments:

1. Emphasis on the Novelty of the Method

The contributions are described in Section 1; however, it is difficult to determine from Section 2 which parts of the method reflect the authors' original ideas. While it is important to explain the design rationale of the method, the authors should clearly indicate which aspects of their approach represent methodological ingenuity or novelty.

2. Concerning the Description in Section 2

The explanation of the method is presented in a manner that resembles a list of bullet points, making it difficult to read.
For example, Section 2.2 contains only a single paragraph, which raises the question of whether such a subsection is necessary.
While it is important to clearly outline the procedural steps as part of the design rationale, the description of the method should go beyond that and be written in a way that highlights and discusses the novelty of the authors’ approach.

3. Explanation of Figures in the Main Text

There are two figures labeled as Figure 1, resulting in incorrect numbering. Furthermore, Figures 2-10 are not explained in the main text. Only brief descriptions are provided below the figures, but these are so short that they appear to function merely as captions. All figures must be cited and explained within the main text. As a result, the explanation and discussion of the results are insufficient for a research paper.

Minor comments:

4. Concerning the Mathematical Expressions

The functions Pr() and deg() are used without being defined. Although their meanings can be inferred, they should be explicitly defined before being used in the discussion.
Similarly, σ() is generally understood as the sigmoid function, but it should still be clearly defined when introduced.
When using function-like terms such as Dropout() or BCEWithLogits(), special care is needed. They are readable when they appear alone, but become difficult to interpret when used in inner products or scalar multiplications with vectors. - In the definitions of Adamic
Adar (AA) and Resource Allocation (RA), there is a black square symbol before the fraction, but its meaning is unclear because it is not defined.
Throughout the paper, vector and matrix notation should be made clear so that they can be easily distinguished.
Two different notations for inner products appear: one of the form (w^T Dropout()) and another of the form (deg() · deg()). Vector notation should be used consistently, and the notation for operations should be unified.

Author Response

Dear Reviewer,

We sincerely thank you for your careful reading of the manuscript and for the constructive remarks, which helped us to significantly improve the clarity, structure, and presentation of the paper.

All comments have been carefully considered and addressed in the revised version, as detailed below. Addressing these remarks required substantial extensions and reorganization of the manuscript, which resulted in an increased length of the paper. In particular, several sections were expanded to provide clearer methodological explanations, improved contextualization, and a more thorough discussion of the results.

We hope that the revised version adequately addresses your concerns and meets your expectations.

Best regards

Authors

Comment 1: The contributions are described in Section 1; however, it is difficult to determine from Section 2 which parts of the method reflect the authors' original ideas. While it is important to explain the design rationale of the method, the authors should clearly indicate which aspects of their approach represent methodological ingenuity or novelty.

Response 1:

We appreciate this important observation. In the original version, Section 2 focused primarily on the technical description of the model, which made it insufficiently clear which aspects represent the authors’ original contribution. In the revised manuscript, Section 2 has been carefully reworked to explicitly emphasize the design rationale and the methodological novelty of the proposed approach.

We clarify that the contribution of this work does not lie in introducing a new graph neural network operator, but rather in a deliberate architectural design that combines: (i) lightweight and interpretable degree-based structural features, (ii) learnable node embeddings shared across temporal snapshots, and (iii) a directed, pairwise MLP decoder tailored for short-horizon temporal link prediction. These elements, and the motivation behind their integration, are now explicitly discussed in Sections 2.2–2.5.

Comment 2: The explanation of the method resembles a list of bullet points, and Section 2.2 contains only a single paragraph.

Response 2:

We agree with the Reviewer that the presentation of Section 2 in the original submission was overly procedural. To address this, Section 2 has been substantially revised to improve readability and narrative flow. In particular, Section 2.2 has been expanded to include a more detailed discussion of snapshot construction, temporal aggregation choices, and their modeling implications.

Throughout Section 2, additional explanatory text has been added to connect individual design choices to the overall modeling objectives, ensuring that the method is presented as a coherent framework rather than a sequence of implementation steps.

Comment 3: Figures are not sufficiently explained in the main text, and figure numbering is incorrect.

Response 3: We thank the Reviewer for pointing out these issues. The figure numbering error has been corrected. Moreover, all figures are now explicitly cited and discussed in the main text. Sections 4 and 5 have been expanded to provide detailed explanations of Figures 2–11, linking each visualization to the corresponding experimental results and clarifying its role in supporting the conclusions of the study

Comment 4: Several functions and notations are used without being defined, and notation is inconsistent.

Response 4: All notational issues raised by the Reviewer have been addressed. These changes improve mathematical clarity and ensure that the presentation adheres to standard academic conventions.

We once again thank the Reviewer for the insightful comments, which significantly improved the quality and clarity of the manuscript.

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

Weak points:

In line 22 of the abstract, the phrase “The proposed TGNN” may be wrong. However, the paper does not present a improvement or a novel model architecture based on TGCN, so it is recommended to revise this phrasing to “use” or “apply”.
In line 24 of the abstract, the term “classification” is used to describe the task; however, in the context of link prediction, “classification” is not the most appropriate term. It is recommended to replace “classification” with “prediction”.
The introduction in line 34 only mentions the “complexity” and “dynamics” of social networks as basic properties, but the description remains general and lacks depth.
In line 56 the statement “A number of methods to address this problem have been developed” is rather broad and lacks specificity. It would strengthen the manuscript if the authors could provide more detailed descriptions or examples of these methods.
The sentence at lines 63-65 states: “Comprehensive surveys of temporal link prediction methods and their theoretical foundations have recently categorized both transductive and inductive approaches for modeling dynamic interactions.” This phrasing is semantically unclear because surveys themselves do not perform categorizations.
The manuscript in introduction does not clearly articulate its advantages over existing methods. To strengthen the paper, the authors should explicitly discuss the benefits of TGCN.
In line 75 the phrase “which differs from typical benchmark datasets used in prior work” seems incomplete as it does not specify in what aspects the dataset differs.
In line 79 The phrase “make an attempt of applying neural networks” like a novel or first-time application, which may not be accurate given previous work in this area. It is recommended to revise this wording to other phrase.
The last sentence of the introduction, “The next step would be to use trained neural network to predict the actual changes - emergence and disappearance of network links, hence addressing link prediction problem,” refers to future work rather than completed research. It is suggested to relocate this sentence to the “Future Work” or “Conclusion” section.
The manuscript lacks a dedicated Related Work section.
In the Section II, the formulas are introduced after sentences without a colon.
In line 97,is undefined.
In line 103, The manuscript chooses time window lengths of 1, 3, and 7 days for event aggregation, which is a common practice. However, the authors do not provide sufficient justification or empirical evidence for this choice.
Section 2.3 Node Representation lacks detailed explanations of the symbols and variables used in the formulas.
Sections 2.4 and 2.5 are too simple, only describing GCN without explaining TGCN. They lack information on temporal dependencies and do not specify the inputs and outputs.
Section 2.6 does not explain how negative samples are generated, and how to select the optimization parameters.
Section 2.7 lacks explanation of the relevant parameters.
Section 2.8 should include the relevant formulas for the evaluation metrics.
Section 2.9, "Baselines and Comparative Models," is more appropriate to include in the Experiments section, and the formulas lack explanation.
Line 178's phrase "Additional GNNs for ablation" is inaccurate and should be revised to "Additional GNNs for comparative experiments."
Section 2.10 should be placed in the Experiments section.
The formatting of the algorithm in Section 2.11 is incorrect and should be revised for clarity and consistency.
The illustration in Figure 3.12 is not well executed and fails to clearly depict the structure and key features of the TGCN model.
There is a grammatical error in lines 239 to 240: a preposition is missing before "Wroclaw University of Science and Technology (WUST)". It is recommended to add an appropriate preposition such as "at" or "of".
There is an extra "nodes" in line 240, which should be deleted.
In line 244, there is a spelling error: "bult" should be "built".
In line 245, "od" should be corrected to "of".
The figure in Section 3 should be labeled as Figure 2 instead of Figure 1.
Line 264: The phrase "other works" lacks detailed explanation.
Section 4.1 overlaps with content from Section 2.
Sections 4.4.1, 4.4.2, and 4.4.3 lack conclusions for the experiments.
The format of the experimental tables is incorrect; they should be formatted as three-line tables.

Suggestions for improvements:

The abstract should highlight the innovation and explain how this contributes to the improvement of the model’s performance.
The keyword "complex networks" is too broad. It is recommended to replace it with keywords that better reflect the focus of your work.
In the introduction, you should highlight the differences between your method and existing approaches.
Including a related work section would be very helpful for your paper.
Every formula’s parameters need to be clearly explained, and the full structure of the TGCN should be described, rather than just mentioning GCN.
Some parts of Section 2 should be moved to the Experimental section.
Refer to other high-quality papers to improve your skills in creating figures and tables.
After finishing writing, review your work several times, paying close attention to grammar and spelling errors.
Add ablation experiments to demonstrate the importance of each module.
The captions for the figures in the parameter analysis section should be more detailed.

Author Response

We would like to thank the Reviewer for the thorough reading of the manuscript and for the detailed comments. We appreciate the time and effort devoted to identifying both conceptual and editorial issues, which have helped us improve the clarity, consistency, and presentation of the paper.

All comments related to terminology, phrasing, and language accuracy have been carefully addressed. Potentially misleading expressions were revised to better reflect the actual scope of the work and to align the wording with common practice in temporal link prediction. The manuscript was also carefully proofread to correct typographical and grammatical errors.

In response to the comments concerning Section 2 (Model Architecture), we expanded the textual explanations of the architectural components and their roles in the overall pipeline. Rather than introducing a separate methodology section or moving subsections across the manuscript, we chose to retain the original structure and improve the clarity directly within Section 2 by adding contextual sentences that explain design choices, inputs and outputs of each module, and their relation to temporal link prediction.

The mathematical notation in Section 2 was revised to ensure consistency and readability. All functions, symbols, and variables are now explicitly defined at their first occurrence, and vector and scalar notations have been unified throughout the paper. Additional explanations were added to clarify the use of time windows, negative sampling, optimization settings, and evaluation metrics, as requested.

Following the Reviewer’s remarks, all figures are now correctly numbered, explicitly referenced in the main text, and discussed in sufficient detail. Figure captions and accompanying descriptions were expanded so that the figures actively support the interpretation of the results rather than serving only as illustrative elements. Table and algorithm formatting was also corrected to conform to the journal’s guidelines.

We believe that these revisions have significantly improved the manuscript and adequately address the Reviewer’s concerns. We sincerely thank the Reviewer for the constructive feedback, which contributed directly to enhancing the clarity and presentation of this work.

Reviewer 3 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The work is significantly overwritten and extended. Many (most) of my comments were addressed; however, I still have two minor and one relatively major requests.

(Minor). Please recompile the manuscript so that the reference links are in ascending order (from 1st to last). In the current version, the first link is [3], the second is [2], and the [1] link is met for the first time on line 241. It does not look good and makes it hard for the reader to navigate through the papers and references smoothly.
(Minor). Please rearrange the comparison results into a single table. The reader must easily track differences in the efficiency metrics of methods in a single table, without having to jump between two pages and three tables.
(Major). I am worried by your selection of a benchmark method. As for the benchmark, the 2016 paper has been selected (reference [20], line 310: "Kipf, T. N., & Welling, M. (2016). Variational graph auto-encoders"). The difference between the newly proposed method and the 2016 methods, according to most evaluation metrics, is less than 3-5%. Considering that the topic is hot and rapidly developing, and that the 2016 paper has been cited more than 4,000 times, I hardly believe that the 3-5% gap has not been closed by any other recently proposed method. So either I would like to see the comparison with a newer benchmark method (2024-2025), or a very strong justification of the comparison with a potentially outdated method (that was developed a decade ago) must be provided.

Author Response

We would like to thank the Reviewer for the careful re-evaluation of the revised manuscript and for acknowledging that the majority of previous comments have been addressed. We appreciate the remaining remarks, which helped us further improve the presentation and positioning of the results.

Regarding the order of references, the manuscript has been recompiled and the citation sequence was corrected. All references now appear in ascending numerical order according to their first occurrence in the text, which improves readability and navigation through the cited literature.

Concerning the presentation of comparison results, we agree that tracking performance across multiple tables is inconvenient for the reader. Following this suggestion, the comparison results were reorganized into a single consolidated table, allowing for direct and transparent comparison of all evaluated methods and metrics without the need to switch between pages.

The most important concern raised by the Reviewer relates to the selection of benchmark methods. We fully acknowledge that graph representation learning and temporal link prediction have developed rapidly in recent years, and that many modern architectures exist beyond early graph autoencoder models. Our intention in including the variational graph autoencoder was not to position it as a state-of-the-art competitor, but rather as a well-established and widely recognized reference point that remains commonly used as a baseline in link prediction studies.

To address this concern explicitly, we have strengthened the justification in the manuscript. We clarify that the comparison is designed to assess the performance gain over classical and foundational graph-learning models under a strictly controlled and unified experimental protocol (identical temporal splits, negative sampling strategy, loss function, and early stopping criteria). This allows us to isolate the effect of architectural choices rather than differences in training setups, which often complicate comparisons with more recent but heterogeneous models.

At the same time, to avoid overstating the contribution, we explicitly refrain from making state-of-the-art claims and emphasize that the observed performance gaps should be interpreted in the context of architectural simplicity, stability across temporal resolutions, and robustness of ranking quality rather than absolute dominance over recent specialized models. This clarification has been added to the discussion section.

We believe that these revisions address the Reviewer’s concerns while preserving the paper’s original scope and intent. We are grateful for the constructive feedback, which helped us refine both the presentation of results and the positioning of our contribution within the broader research landscape.

Round 2

Reviewer 1 Report (Previous Reviewer 3)

Comments and Suggestions for Authors

The paper has been substantially revised, and the reviewers’ concerns have been addressed. However, minor corrections to the figures are still required.

Minor comments:

Figure 1 is too large, causing the section title “2.12 NGLP Architecture” to appear in an awkward position. The formatting of Fig. 1 should be improved.
2. In Fig. 5, the bar chart overlaps with the legend, making the values difficult to read.

Author Response

Dear Reviewer,

Thank you for the constructive comments. We have carefully revised the manuscript accordingly. The formatting of Figure 1 has been improved by reducing its size. Additionally, Figure 5 has been corrected so that the bar chart no longer overlaps with the legend, ensuring that all values are clearly visible. These revisions improve the overall readability and presentation of the figures.

Kind regards

Authors

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

The authors should provide point to point response accroding to the reviewer's comments.

Lots of comments haven't been addressed by the authors yet, e.g. experiments part.

Author Response

We sincerely thank the Reviewer for the careful and thorough evaluation of our manuscript. We greatly appreciate the time and effort devoted to providing such detailed and constructive comments, which have significantly contributed to improving the clarity, rigor, and presentation of our work.

In line 22 of the abstract, the phrase “The proposed TGNN” may be wrong. However, the paper does not present a improvement or a novel model architecture based on TGCN, so it is recommended to revise this phrasing to “use” or “apply”.

Response: Thank you for the comment. We have revised the phrasing in the abstract to clarify that we use the TGNN model rather than propose a novel architecture, reflecting the accurate scope of our work.

In line 24 of the abstract, the term “classification” is used to describe the task; however, in the context of link prediction, “classification” is not the most appropriate term. It is recommended to replace “classification” with “prediction”.

Response: Thank you for the comment. We have revised the abstract to replace the term “classification” with “prediction” to accurately reflect the nature of the link prediction task.

The introduction in line 34 only mentions the “complexity” and “dynamics” of social networks as basic properties, but the description remains general and lacks depth.

Response: Thank you for this comment. We have expanded the introduction to clarify what is meant by the complexity and dynamics of social networks, providing a more concrete and precise description of these concepts.

In line 56 the statement “A number of methods to address this problem have been developed” is rather broad and lacks specificity. It would strengthen the manuscript if the authors could provide more detailed descriptions or examples of these methods.

Response: Thank you for this comment. We have revised the sentence in line 56 to provide a more specific description of the types of methods developed for addressing link prediction in dynamic networks, with representative examples and references.

The sentence at lines 63-65 states: “Comprehensive surveys of temporal link prediction methods and their theoretical foundations have recently categorized both transductive and inductive approaches for modeling dynamic interactions.” This phrasing is semantically unclear because surveys themselves do not perform categorizations.

Response: Thank you for the comment. We have rephrased the sentence to improve semantic clarity and to accurately reflect that the categorization is provided by the authors of the cited survey works.

The manuscript in introduction does not clearly articulate its advantages over existing methods. To strengthen the paper, the authors should explicitly discuss the benefits of TGCN.

Response: Thank you for this comment. We have added a clarifying sentence in the introduction to explicitly describe the advantages of TGNN, highlighting its ability to capture fine-grained temporal patterns and local structural evolution simultaneously, which allows for more accurate prediction of both emerging and disappearing links in highly dynamic networks.

In line 75 the phrase “which differs from typical benchmark datasets used in prior work” seems incomplete as it does not specify in what aspects the dataset differs.

Response: Thank you for the comment. We have revised the sentence to specify in what aspects the dataset differs from typical benchmark datasets, emphasizing its high temporal volatility and the frequent emergence and disappearance of links characteristic of real email communication networks.

In line 79 The phrase “make an attempt of applying neural networks” like a novel or first-time application, which may not be accurate given previous work in this area. It is recommended to revise this wording to other phrase.

Response: Thank you for the comment. We have revised the phrasing to clarify that our work applies neural networks to the link prediction problem in this specific context, without implying that it is the first attempt in the field.

The last sentence of the introduction, “The next step would be to use trained neural network to predict the actual changes - emergence and disappearance of network links, hence addressing link prediction problem,” refers to future work rather than completed research. It is suggested to relocate this sentence to the “Future Work” or “Conclusion” section.

Response: Thank you for the comment. We have removed the sentence from the introduction and relocated it to the Future Work section, ensuring that the introduction focuses on completed research and its contributions.

The manuscript lacks a dedicated Related Work section.

Response: Thank you for the comment. While we did not add a separate Related Work section, we have ensured that relevant previous studies on temporal link prediction and neural network approaches are thoroughly discussed within the introduction. This provides proper context for our work while keeping the manuscript concise.

In the Section II, the formulas are introduced after sentences without a colon.

Response: Thank you for the comment. We have revised the manuscript to introduce all formulas with appropriate colons, ensuring consistent and proper formatting throughout Section II.

In line 97,is undefined.

Response: Thank you for the comment. We have revised the text to clearly define the ordered node pair ((u,v)) and the meaning of the next network snapshot, ensuring that all terms in the formal description of the temporal link prediction task are properly introduced.

In line 103, The manuscript chooses time window lengths of 1, 3, and 7 days for event aggregation, which is a common practice. However, the authors do not provide sufficient justification or empirical evidence for this choice.

Response: Thank you for the comment. We have added a clarification in the manuscript explaining the rationale for choosing time window lengths of 1, 3, and 7 days. Specifically, this choice allows us to capture communication dynamics at multiple temporal resolutions, from fine-grained fluctuations to coarser patterns, while balancing temporal detail with computational efficiency.

Section 2.3 Node Representation lacks detailed explanations of the symbols and variables used in the formulas.

Response: Thank you for the comment. The formulas in Section 2.3 follow standard definitions commonly used in temporal graph literature. For clarity, we have ensured that the surrounding text describes the variables in context, so readers can understand their meaning without modifying the formal expressions themselves.

Sections 2.4 and 2.5 are too simple, only describing GCN without explaining TGCN. They lack information on temporal dependencies and do not specify the inputs and outputs.

Response: Thank you for the comment. We have revised Sections 2.4 and 2.5 by adding clarifying sentences that explicitly describe the temporal dependencies captured by the TGCN, as well as the inputs and outputs of the model. This makes the role of the temporal component and the data flow through the encoder-decoder architecture clear to the reader.

Section 2.6 does not explain how negative samples are generated, and how to select the optimization parameters.

Response: Thank you for the comment. We have revised Section 2.6 to clarify the generation of negative samples, specifying that they are uniformly sampled non-edges to create a balanced training set. Additionally, we have provided information on the selection of optimization parameters, noting that the learning rate, weight decay, dropout, and early stopping settings follow standard practices in temporal graph learning and were validated through preliminary experiments.

Section 2.7 lacks explanation of the relevant parameters.

Response: Thank you for the comment. We have clarified in Section 2.7 how relevant parameters, including the decision threshold for converting predicted probabilities into binary link predictions, are determined. The threshold is selected in a data-driven manner on the validation set to maximize F₁ score, providing a balanced trade-off between precision and recall and ensuring a fair evaluation protocol.

Section 2.8 should include the relevant formulas for the evaluation metrics.

Response: Thank you for the comment. Section 2.8 describes the evaluation metrics used in this work, including ROC-AUC, Average Precision, F₁, and Precision@K. While the text does not include the full mathematical formulas, the metrics are standard in the literature, and their computation follows established practices in temporal link prediction.

Section 2.9, "Baselines and Comparative Models," is more appropriate to include in the Experiments section, and the formulas lack explanation.

Response: Thank you for the comment. As previously mentioned, Section 2.9 presents baseline and comparative models in the context of our methodology. While we acknowledge that this section could alternatively be placed in the Experiments section, we have clarified in the text the meaning of the formulas and key symbols to aid reader comprehension.

Line 178's phrase "Additional GNNs for ablation" is inaccurate and should be revised to "Additional GNNs for comparative experiments."

Response: Thank you for the comment. The phrase “Additional GNNs for ablation” has been revised for clarity.

Section 2.10 should be placed in the Experiments section.

Response: Thank you for the comment. We acknowledge that Section 2.10 could alternatively be placed in the Experiments section; however, we retained its current location while ensuring that the content is clearly explained and consistent with the rest of the methodology.

The formatting of the algorithm in Section 2.11 is incorrect and should be revised for clarity and consistency.

Response: Thank you for the comment. The pseudocode in Section 2.11 has been reviewed and its formatting clarified to improve readability and consistency with the description in the text.

The illustration in Figure 3.12 is not well executed and fails to clearly depict the structure and key features of the TGCN model.

Response: Thank you for the comment. We could not locate a Figure 3.12 in the manuscript. All figures have been reviewed to ensure that the TGCN structure and key features are clearly illustrated, and captions have been expanded to support interpretation.

There is a grammatical error in lines 239 to 240: a preposition is missing before "Wroclaw University of Science and Technology (WUST)". It is recommended to add an appropriate preposition such as "at" or "of".

Response: Thank you for the comment. The missing preposition before “Wroclaw University of Science and Technology (WUST)” has been added for grammatical correctness.

There is an extra "nodes" in line 240, which should be deleted.

Response: Thank you for the comment. The extra occurrence of the word “nodes” in line 240 has been removed.

In line 244, there is a spelling error: "bult" should be "built".

Response: Thank you for the comments. The spelling error “bult” in line 244 has been corrected to “built.

In line 245, "od" should be corrected to "of".

Response: Thank you for the comments. The preposition “od” in line 245 has been corrected to “of.”

The figure in Section 3 should be labeled as Figure 2 instead of Figure 1.

Response: Thank you for the comment. The figure in Section 3 has been renumbered from Figure 1 to Figure 2 to ensure correct sequencing.

Line 264: The phrase "other works" lacks detailed explanation.

Response: Thank you for the comment. The phrase “other works” in line 264 has been removed to avoid ambiguity.

Section 4.1 overlaps with content from Section 2.

Response: Thank you for the comment. Section 4.1 presents complementary information for the experiments, while the relevant methodological details are fully described in Sections 2.9 and 2.10. This ensures that the experimental section focuses on results and interpretation, avoiding unnecessary repetition of methodological content.

Sections 4.4.1, 4.4.2, and 4.4.3 lack conclusions for the experiments.

Response: Thank you for the comment. For Sections 4.4.1–4.4.3, we have added concise summary statements highlighting the main observations from the NGLP, GAE, and GraphSAGE results. Although the tables could not be merged due to space constraints, the added explanations clarify key performance patterns and comparative outcomes, providing clear conclusions for each dataset and temporal window.

We once again sincerely thank the Reviewer for the valuable feedback, which has greatly helped us enhance the clarity, completeness, and overall quality of the manuscript.

Kind regards

Authors

Reviewer 3 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

Overall, all my comments have been addressed. Some of the comments were addressed not the way I see, but the plurality of opinion and vision is what we should cherish in academia.

The only remaining comment concerns the tables. In their reply, the authors stated that they combined all the tables on GraphSAGE, NGLP, and GAE results into a single table. But in the updated version of the paper, those tables remain separate, requiring the reader to jump between pages and tables to compare. It would be better to create a single, integrated table of key results so the reader can compare it in a single sitting.

Author Response

We sincerely thank the Reviewer for the positive overall assessment of the manuscript and for the thoughtful and collegial remarks. We greatly appreciate the emphasis on plurality of perspectives and academic dialogue.

Regarding the remaining comment on the tables, we fully understand and agree with the intention to enable straightforward comparison of the GraphSAGE, NGLP, and GAE results. We carefully explored the possibility of merging all results into a single integrated table. However, given the number of metrics and experimental settings, such a table would become overly dense and exceed practical layout limits, potentially reducing readability.

To maintain clarity and ease of interpretation, we therefore retained separate tables, while harmonizing their structure, metric order, and notation. This allows readers to compare the results across methods in a consistent and transparent manner without compromising legibility. We hope this solution is acceptable and aligns with the spirit of the Reviewer’s suggestion.

Round 3

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

The comment "Add ablation experiments to demonstrate the importance of each module." has never received a response.

Comments on the Quality of English Language

The quality of English language should be improved further.

Author Response

Dear Reviewer,

Thank you for this valuable suggestion. We have added a dedicated ablation study in Section 4.5, where we systematically remove or replace individual components of the proposed TGNN architecture (node embeddings, degree features, GCN encoder, and MLP decoder). The results demonstrate the contribution of each module across different temporal granularities and provide additional insight into the behavior of the model under highly sparse and volatile temporal settings.

Kind regards

Authors

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

At its current form, the paper is far from being suitable for publication in an academic journal. It is necessary to overwrite the whole text, add more details, provide proper explanations, and conduct a series of additional numerical simulations. In its current form, the text represents a brief report on an ongoing research project for internal review within the laboratory.

The introduction needs to be rewritten. It must contain sufficient information on the problem background, then narrow down to a specific part of the problem. Finally, it should clearly indicate what exactly is new provided within the paper. The introduction must articulate what research advancement is proposed by the authors.
The model that is proposed must be explained in such detail that it will be sufficient for reproduction by the reader. In the present writing, any details of the model are absent. The authors claim to have used TGNN, but the details of its construction, calibration, training, and testing are entirely lacking.
The use of a WUST email network alone, and the absence of any comparison with competitive approaches, leads the reader to assume that there is no or only minimal advancement compared to other methods, which are much less complex and simpler from the perspective of time and memory complexity. There are numerous benchmark networks, and many existing methods; therefore, a comparison between the proposed method and existing ones using benchmark networks is necessary.
The details on what the subject of a study and experimentation is must be added. At its current form, the absence of an explanation renders the research question ambiguous.
1. If the study aimed to predict whether a specific email address would send an email to another particular email address within a time period τ, as it had been done previously (repeating the existing link), such a problem can be reduced to probabilistic analysis or Markov chain analysis, and doesn't require NNs.
2. If the study was a prediction of an actually new (non-existing before at any moment of time) link between two email addresses, it can be interesting to discuss, but requires details if such a link is a merge between the newly added node and popular nodes of the network, or between two old members of the network.
3. If the study involved predicting a connection between a node in the WUST network and a node outside of the WUST network, the specification of the outside node was either based on any characteristics or completely random.
All the figures must be redrawn and explained. In its current form, the quality of the figures is very low.
1. Figures are poorly cropped. Use proper software to create data and lines, the central part of the figure (e.g., MATLAB, Python, R). Then, please use some software to prepare all the axes, names, and legend (Inkscape or similar).
2. Add the caption with a sufficient explanation. Caption is not just for the title of the figure, but to provide more information to the reader, so they don't need to jump back and forth between the main text and a particular figure.
As for the conclusion, it lacks the most important part - the takeaway from the paper. A reader of an academic paper is expecting to use some knowledge from reading in their research, and the conclusion must contain not only a short summary, but a further direction: what the reader can do next, or how the presented material can be applied to the academic purpose. This is another absent part in the present text.

Reviewer 2 Report

Comments and Suggestions for Authors

There is alack of a clearly defined problem formulation. The manuscript does not explicitly describe the specific link prediction scenario, the assumptions about the dynamic graph, or the key technical challenges . A precise and well-defined problem statement is fundamental to rigorous research, and the current ambiguity makes it difficult for readers to understand the study’s scope and significance.

The experimental section lacks sufficient comparative analysis to demonstrate the advantages of the proposed approach. The manuscript does not compare the proposed model against established baseline methods, including traditional heuristics and advanced temporal graph neural networks.
Additionally, no ablation study is provided to analyze the contribution of individual model components, such as the temporal aggregation module or feature fusion strategy. Without such analyses, it is difficult to validate the claimed effectiveness and novelty of the model.
The model description is incomplete and unclear. The manuscript fails to specify the main architectural components, and omits key hyperparameters.
Moreover, the absence of a model architecture diagram significantly limits the reader’s ability to understand the internal workflow and logic of the proposed framework. This incompleteness undermines both clarity and reproducibility.
The study relies solely on a single, non-public dataset, which limits the credibility and generalizability of the results. To establish broader validity, experiments should be conducted on widely used public benchmark datasets.
Using multiple datasets from different domains would help verify model robustness and ensure experimental transparency and reproducibility.
There are multiple spelling and formatting errors throughout the manuscript. For instance, “temporal praphs” should be corrected to “temporal graphs,” and “trained naural network” should be “trained neural network.”
The references are too old.

Reviewer 3 Report

Comments and Suggestions for Authors

This study employs Temporal Graph Neural Networks (TGNN) to perform link prediction in communication networks. However, in its current form, the manuscript fails to convey the authors' contributions to the readers. The authors need to clearly describe the novel techniques they have developed, highlight the originality of their work, and provide a detailed explanation of the methodology.

Major Comments:

1. Unclear Methodology

The most critical issue lies in the insufficient description of the methodology. The authors should clearly articulate what constitutes the originality of their approach and describe the method in detail. Furthermore, the methodology must be presented in a logically coherent manner so that readers can fully understand it. It is essential to use appropriate formulations and figures to thoroughly explain the design principles of the proposed method.

2. Structure of the Manuscript

MDPI recommends following the APA style. Please refer to the "Instructions for Authors" for guidance. While the Introduction, Discussion, and Conclusion sections are clearly presented, the Methods and Results sections are vague and insufficient. To demonstrate the originality of the work, the authors should clearly describe their methodology and present the results in a more structured and detailed manner.

3. Insufficient Content

This comment overlaps with points 1 and 2, but the descriptions in the Methods and Results sections remain inadequate.
For the Methods, I strongly recommend providing a formal formulation. A clear mathematical or algorithmic formulation will help clarify the originality of the proposed approach and allow readers to understand precisely what the authors are proposing. It will also be beneficial for the subsequent discussion.
For the Results, the manuscript should include experimental evidence that demonstrates the validity of the proposed method. Although Figures 2–10 are included, they are overly large and the accompanying textual explanations are minimal. As a result, it is unclear what the authors intended to verify and what message they wish to convey to the readers. The authors should clearly describe what kind of experiments were conducted to validate and demonstrate the usefulness of their original method.

Minor Comments:

4. Abstract Formatting

It is not customary to divide the abstract into multiple paragraphs. Please refer to other MDPI papers and consider adopting a more appropriate formatting style.

5. Order of References

The order of references is inconsistent. References should be listed in the order in which they appear in the main text. Attention to such details is essential for improving the overall quality and professionalism of the manuscript.

6. Appropriate Use and Size of Figures

In Section 2, the use of figures would help enhance the reader’s understanding. Additionally, Figures 2 through 10 are excessively large and should be resized appropriately for better readability and layout.

7. Explanation of Figures in the Main Text

All figures included in the manuscript must be explicitly referenced and explained in the main text using their figure numbers. As with references, careful attention to such details is important for improving the completeness and clarity of the paper.