On the Relationship Between Relevance and Conflict in Online Social Link Recommendations

Summary

The paper analyzes the relationship between relevance and opinion conflict/disagreement on online social media platforms, particularly in link recommendations. Using the Friedskin-Johnsen model that models opinion dynamics, the paper shows how link additions change online conflicts and characterize the gap in conflict reduction, depending on the reduced link recommendation. The paper evaluates their approach using real-world datasets, obtained from previous work, from Twitter and Reddit.

Strengths

The paper focuses on an important and timely problem related to online polarization. I liked that the paper presents all the necessary background information about their models, as well as the assumptions that they made. Also, I believe that evaluating the proposed approach on multiple social networks and multiple recommendation algorithms makes the paper and its findings more robust.

Weaknesses

I have a couple of concerns with this paper. First, the formulation of the social network as an undirected graph is not very representative of what is actually happening on popular social networks like Twitter. In reality, the social network is undirected, which raises the questions on how the presented results might change when considering the fact that the graph is undirected instead of directed. I suggest to the authors to expand their discussion on this point in the paper. Perhaps more importantly, the paper overlooks a substantial shift that happened in social media platforms. Recently, we are observing the use of AI recommendation systems that are used for recommending and delivering content to the end-users. That is, people are exposed to content mainly based on the recommendation algorithm and, to a lesser extent, to content that is coming from people that a user follows or are friends with (e.g., see “For You” algorithmic feeds in Twitter, TikTok, etc.) These recommendation algorithms are becoming an integral part of most popular social networks nowadays. Given this, I am wondering whether the proposed approach can even be applied on this use case or how the presented results change when considering that there is an algorithm that determines what each user sees. I suggest to the authors to discuss this aspect and whether the proposed approach can be leveraged to study the same problem under these considerations.

Questions

1. Can the proposed approach be applied to directed graphs? If yes, what was the rationale for using undirected graphs in this work? 2. Can the proposed approach be applied when considering that there is a separate entity (i.e., recommendation algorithm) responsible for the content delivery to users?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have addressed some limitations of their work. However, I strongly encourage them to have a look at the weaknesses part of this review and expand their limitations part with the two points that are described.

Summary

This paper motivates that previous works have studied the two important aspects providing relevance and reducing conflict mainly in isolation, thus aims to explore the relationship between them. The paper uses FJ model from opinion dynamics as base model to first derive a closed form expression for quantifying changes of conflict, which is then used to analyze the conflict reduction effect caused by link addition. Based on the analysis, the paper then characterizes the conflict-reducing links and how these align with relevant links. To put the analysis in practice, a metric termed conflict-awareness is proposed to quantify the alignment between relevance and conflict minimization. The proposed metric is applied to various link recommendation algorithms on two real-world social networks. The findings show that some more accurate algorithms are indeed better at reducing conflict.

Strengths

* The idea of analyzing the trade-off between relevancy and conflict in link prediction using Friedkin-Johnsen model is interesting and new to my knowledge. * Design choices of the presented analysis are properly justified. * The paper is easy to follow.

Weaknesses

* The main claim in abstract "To this date, however, we have very little understanding of how these two implications of link formation relate to each other..." is not fully justified. A 2021 PNAS paper titled "Link recommendation algorithms and dynamics of polarization in online social networks" also tackle the issue of polarization and conflict in online social networks, specifically exploring the role that link recommendation algorithms play. Without establish the link to this 2021 paper as well as the papers mentioned therein, the main claim appeals unconvincing.

Questions

* How would the paper been positioned given the PNAS 2021 paper? * About the limitation of the proposed approach, I am wondering whether any observation can be made by dropping the equilibrium assumption of the Friedkin-Johnsen model? It would be interesting to analyze the conflict in the network form the time point of adding a link until the equilibrium is reached. * F&J model assumes the aggregated opinion would be assimilated internally, but in some cases when introducing two opinionated nodes to each other, backfire could happen. Would this type of effect complicates the analysis? or a more vague question, how would the proposed model align with the logged events / conflict proxy measures in real world data?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

* The authors pointed out two constraints: a paradox that reducing conflict could potentially stress users, and the choice of using directed versus undirected graphs for studying opinion dynamics, both limitations arising from the use of the FJ model.

Summary

With the flourishing of online social media and social networks, the effects of polarization and disagreement brought about by link recommendation methods have become increasingly pronounced. Currently, a more widespread view is the filter bubble theory. But the empirical evidence to support him is limited, and there is a lack of detailed understanding of the strength that the filter bubble effect caused by social recommendations. To deal with these issues, this paper explores the theoretical evidence and principled features of relevance and conflict in online social link recommendation. First, the paper investigates the amount of change in opinion conflict caused by adding general links and concludes that purely adding social links does not increase opinion conflict, also it proposes two criteria for finding conflict-minimizing links in social networks; then the paper introduced Conflict awareness to evaluate the ability of a link recommendation model to reduce conflicts; finally, it discussed the limitations of the above theoretical analysis based on the FJ model.

Strengths

The main strengths of this paper are the following: 1. originality This paper explores theoretical evidence and principled features of the effects of polarization and disagreement brought about by link recommendation algorithms in social networks. It also provides the first analysis of the question of whether the most likely accepted links in social networks are fundamentally different from those that reduce conflict by using the Friedkin-johnsen model of opinion dynamics. This paper is of good originality. 2. quality This paper has a clear research question, rich theoretical analysis, solid experiments, and analysis of the limitations of the method, which is of high quality. 3. clarity The paper is clear, logical and well organized. 4. significance This paper is the first to examine whether the most likely accepted links in social networks are different from those that reduce conflicts, and analyzes the relationship between relevance and opinion conflict in link recommendation, challenging the idea that the two are incompatible.

Weaknesses

The theoretical analysis in this paper is rich, the experiments may be limited in space, and it is limited in presentation, but it is necessary to introduce the corresponding experimental detail and experimental settings, please increase the representation.

Questions

This paper analyzes the relationship between relevance and opinion conflict in online social link recommendation and offers a different view on its utterly incompatible perspective. This paper has the following problems: 1. the title of 2.1 is more appropriately changed to Social Network Model; 2. the text in Figure 2 should be labelled for the lines, but the second small figure in Figure 2 is confusingly labelled; 3. The theoretical part of this paper is very rich, but the space of the experimental part is limited, the configuration of the experimental environment is not explained, and only the experiments on conflict awareness are given. Is it more persuasive to put Appendix D5 in the experimental part of the main text? 4.In the conclusion of p241-243, does this distance refer to a specific distance calculation or does it refer generally to a universal distance (Are other distances applicable？). 5. The formulae should be unified and standardized, and the formulae in limitations are not labelled. 6. The number of references is too many, and the number of references in the last five years is few.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

This paper adequately acknowledged the limitations, with a detailed description of the limitations

Summary

Authors theoretically analyze the conflict in the social network under Friedkin-Johnsen model. Those theoretical analysis demonstrates that the addition of links reduces the conflict. In addition, the network meaning of links to minimize conflicts is presented with theoretical support. Finally, based on the conflict-aware score, authors analyze several existing link recommendation algorithms and show that the relevant link recommendation is not fully introducing conflicts.

Strengths

- Thorough analysis has been made with implying the insightful explanation on the social behaviors. - Well-defined conflict-related measurements enable the theoretical analysis as well as the empirical measurements from real-world algorithms and datasets. - Authors propose the conflict awareness that can represents the level of conflict that a given link recommendation algorithm introduces.

Weaknesses

- The claim that the addition of links decreases the conflict is somehow obvious by the opinion formation design. By the mechanism that a give node's opinion is formed by smoothing neighbors' opinions, the addition of links implies more chances of smoothing more opinions.The given model does not have the mechanism of opinions going extreme such as taking the max votes from neighbors. It would be interesting to see if the similar results hold if the aggregation of neighbor opinions is different.

Questions

- In reality, relevance may affect the opinion formation. For example, in Eq (1), a_ij can be correlated with the relevance score. It would be great to see how Section 5 analysis is affected when we incorporate this. - Figure 2 shows downward plots in general. What would be the interpretation?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Authors address that the argument is under the FJ opinion formation model. Also, authors introduce the paradox of conflict and unhappiness, which cannot be analyzed by the manuscript.