You are currently viewing a new version of our website. To view the old version click .
MDPIMDPI
Search all articles
by
  • Yong Han1,
  • Zhenqiao Liu2,* and
  • Hongying Bai3
  • et al.

Reviewer 1: Anonymous Reviewer 2: Anonymous Reviewer 3: Anonymous

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper provides a comprehensive approach for multimodal online public opinion event extraction and trend prediction. This work specifically addresses edible agricultural product safety which is an important factor. For prediction purposes, the work combines multimodal feature fusion, HDBSCAN clustering for event discovery, and a advanced LSTM-PPO deep reinforcement learning model. The methodology is well-structured. All experiments show improvements over several baseline models,

 

In this paper, experiments are conducted on dataset of 3,847 news items which seems very limited. It is suggested to expand the dataset to include a larger volume and a wider variety of sources such as social media posts, forums. This can strengthen the generalizability and robustness of the findings and improve the paper.

 

It is suggested to include pre-trained multimodal architectures. This will provide a more rigorous benchmark and better situate the claimed improvements.

 

The evaluation of the HDBSCAN event discovery in this paper focuses on clustering metrics . I think a qualitative or quantitative analysis of the semantic coherence of the discovered event clusters might provide stronger evidence for its practical utility beyond just numerical cohesion.

 

Some statements are ambiguous. For instance, the contribution of the PPO component versus the LSTM backbone is not well explained. To quantify the specific value added by the reinforcement learning framework, an ablation study comparing LSTM-PPO with a standard LSTM or LSTM-Attention will be interesting to discuss.

 

The computational cost and inference time  of the proposed pipeline is hard to understand, more discussion is needed.

Writer a paragraph in intro to orient the reader about like what is covered in subsequent sections.

There are so many tables and some seems unnecessary i.e.,  Table 2.

 Table 14 spreads over 2 pages. Similarly some sections start at inappropriate page, see “Related work”, why at the bottom of page 2?

Why page 21 is blank?

 

 Many public opinion monitoring requires timely responses and hence  analysis of the model's efficiency and potential bottlenecks is suggested to study its practicality to real  world cases.

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors

Your paper presents a method for predicting public opinion trends using deep reinforcement learning, which improves the model's cognitive capabilities by integrating event feature representation with reasoning techniques. 

Here is my review.

1.- You should improve the abstract, it is long, be more accurate.

2.- The Related Work is well, just improve the limitations of the others researches.

3.- You have posted your contributions, but there are not the information about this. For example, you say that employ a neuronal network, however in the paper is not enough detail about it. The third contribution, you are talking about the construction of a model, in the paper is the comparison of models and not construction. 

4.- It is not clear what are you looking for, you are talking about prediction, but in the paper there are not about that (type, positive, negative, etc), you have compared models, you have indicated what is the best. For example, you were predicted about the opinion of the quality of the products on certain time oriented for marketing issues. Probably is better use or explain more about data mining criteria. 

5.- Less text and more statistics and results graphs. The rare important information, but you have to present it differently. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors
  1. In the paper, the AHP Weighting Justification is needed as the public opinion risk indicator system using the Analytic Hierarchy Process (AHP) does not clearly explain how experts scored their input, why the 1-9 scale was chosen, or how the consistency ratio (CR) was determined. While the weights are provided, there is insufficient validation of the judgment matrices.
  2. The dynamic gating unit is described using formulas (1) to (7), yet it is unclear how it learns to optimize itself and how this impacts the final feature vector.
  3. In the paper although the analysis shows that HDBSCAN is superior, it does not clearly explain how it differs from DBSCAN, except for it being density-adaptive. It would help to tie these differences more to the challenges posed by "complex structures" and "noise" in the public opinion text data.
  4. The LSTM component in the LSTM-Proximal Policy Optimization (PPO) model is mentioned, but the exact diagram showing how the LSTM fits into the architecture is missing. More clarity is needed on how the LSTM processes the sequence of risk indicators (State S) to produce predictions (Action A).
  5. It is suggested for the authors to mention on how the AI and machine learning (ML) are used to manage security and risks in Internet of Things (IoT) networks and to refer from the paper "Optimized Support Vector Machine Based Fused IoT Network Security Management" presented on Artificial Intelligence for the Internet of Things (AIIoT) is relevant. This reference supports discussions on using AI and machine learning (ML) for managing security and risks in Internet of Things (IoT) networks.
  6. In the paper it is to be mention on the model is optimized using the Honey Badger Algorithm for better cyber threat detection in IoT networks is to be considered in the paper. The paper to be mentioned as suggested for the paper: “A Hybrid Autoencoder and Gated Recurrent Unit Model Optimized by Honey Badger Algorithm for Enhanced Cyber Threat Detection in IoT Networks”. This introduces a Hybrid Autoencoder and Gated Recurrent Unit Model. This  relates closely to the current study's use of a sequential deep learning model for time-series forecasting, highlighting how optimization methods can be combined with GRU/LSTM networks in high-volume data scenarios.
  7. It would be suggested to include an introductory a content in the Methodology section (3.3.2) that discusses how the LSTM model works with the PPO algorithm.
  8. The introduction clearly outlines the problem, discussing social sensitivity, the fast rise of food safety public opinion, and the weaknesses of traditional methods. Clearly define the term Deep Reinforcement Learning (DRL) early in the manuscript, explaining that it merges Deep Learning and Reinforcement Learning, as it is a key part of the methodology.
  9. In the related work review the review covers multimodal data, event extraction, indicator systems, and time-series forecasting effectively. It highlights the trend towards multimodal extraction and recognizes a gap in applying character-level/word-level neural networks to Chinese public opinion.
  10. The transition between the limitations of the Transformer model and the need for a DRL model could be smoother. It would be useful to emphasize that DRL is specifically suited for dynamic, sequential decision-making, an area where time-series models, like RNN, LSTM, and Transformer, tend to face challenges.
  11. The conclusion effectively summarizes the methodology focused on multimodal, DRL-based prediction and highlights key findings regarding improved cognitive and predictive capabilities in crucial, safety-sensitive areas.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

My commnets have been addressed.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors

It is much better.