Modelling the Closed Loop Dynamics Between a Social Media Recommender System and Users' Opinions

arXiv:2507.19792 · cs.SI · Submitted 2025-07-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Modelling the Closed Loop Dynamics Between a Social Media Recommender System and Users' Opinions".

Jane: Here's the gist of the paper. Modelling the Closed Loop Dynamics Between a Social Media Recommender System and Users’ Opinions proposes a mathematical model to study the coupled dynamics of a Recommender System (RS) algorithm and content consumers (users) This paper finds that different opinion distributions are more susceptible to becoming polarised than others, many content stances are ineffective in changing user opinions, and creating viral content is an effective measure in combating polarisation of opinions Model Structure The proposed model describes a socio-technological closed-loop system where a Recommender System (RS) and users mutually influence each other The system consists of three key components: the graph representing the social media environment, the user behaviour and opinion dynamics, and the algorithm governing the RS Users hold an opinion xi(t) ∈

−1, 1: , while content has a fixed stance xcj = 2(n − 1)(j − 1) - 1 (1) The RS generates directed edges (cj, ui) from a single selected piece of content j to user i at each timestep t, forming a time-varying bipartite graph G(t) User Behaviour Dynamics Three key processes are modeled for a user when presented with content: the user’s engagement decision, the user’s watch rate, and the user’s opinion dynamics The engagement decision is determined by payoff functions πi(y j i(t)) dependent on the difference x¯i(t) = xi(t)−xzi(t) The probability of choosing an engagement decision y j i(t) is given by P

y j i(t) = s: = exp⁡ to βiπi(s), which is a softmax function The user watch rate wizi(t)(t) is determined by v(t) = xi(t)xzi(t), with wizi(t)(t) = 1 / (1 + e−γv(t)) The opinion evolution follows the classical Friedkin–Johnsen model, where xi(t + 1) = λi h wizi(t)(t)xzi(t) + 1 − wizi(t)(t) (5) RS Mechanism and Payoff The RS uses user-specific variables A j i(t), B j i(t), and C j(t) to calculate a payoff P j i(t) which determines the recommended content The payoff is defined as P j i(t) = (1 − ω)(A j i(t) + B j i(t)) squared + ωC j(t) (9) The RS then uses the softmax function to determine the probability of recommending content c j at timestep t+1: P

zi(t + 1) = cj: = exp⁡ to αP j i(t)/Pk exp⁡ to αP j i(t) (10) Simulation Results and Findings The Monte Carlo simulations investigate the impact of various parameters, specifically the softmax parameter α, content diversity k, and virality filtering parameters ω and δ When investigating α, smaller values correspond to an RS that aims to be more explorative (recommending a range of content) and larger values correspond to an RS that is more exploitative For the NDIC, the population becomes more polarised and radicalised as α increases, with substantial change beyond α = 16 When k = 2, the RS performs best in maximising engagement and watch rate but leads to a hyper-polarised and radicalised population For 3 ≤ k ≤ 11, the RS reduces polarisation and radicalisation in the population as opinions coalesce around the viral content When ω > 0.5, polarization and radicalisation are reduced substantially because all users are more likely to be recommended the same content Viral content emerged as a consistent phenomenon in the proposed model, its effects highlight the closed loop nature of the system The findings suggest there is a “sweet spot” of moderate content diversity (moderate k) that decreases polarisation and radicalisation A limitation of the model is that our current watch rate function results in a user always having a watch rate of approximately 50% when recommended neutral or close-to-neutral content Another limitation is that our virality-based filtering mechanism considers engagement from all users, and extending the framework to consider multiple topics could be useful for studying the rabbit-hole/spillover effect Supplementary Material Results The supplementary material provides results for homogenous population of λi = 0.2, showing that both NDIC and BDIC are consistently below zero, meaning polarisation and radicalisation were always decreased within the populations across the different α values For k > 9 and k is odd, there are notable decreases in the opinion distribution metrics due to a neutral piece of content going viral The results for BDIC show that the RS impacts an initially polarised population less as each user will likely already favour a particular content stance For k ≥ 10, diversity of content does not have a great effect on the BDIC and that distribution remains polarised The results for NDIC show that when k ∈

11, 101: and the change in metrics has asymptotically diminished Extended Results In the extended results for Research Question Three, the contour plots exploring the relationship between ω, δ and when content goes viral within simulations for a BDIC do not substantially differ from those seen in Fig.

Tom: 8 There are no notable differences between Fig.

Jane: 16 and Fig.

Tom: 8 The results seen in Fig.

Paper summary: Tom: Thinking about the title "Modelling the Closed Loop Dynamics Between a Social Media Recommender System and Users' Opinions," what is the actual big picture they are trying to show us here?

Jane: They’re showing how these two things—the algorithm and the users—aren't just separate parts, but they are deeply coupled in a continuous cycle where each one affects the other over time.

Lu: The implication for AI research is that we need to move beyond studying recommender systems in isolation and start modeling them within this kind of dynamic environment that includes opinion evolution.

Meng: From an engineering standpoint, it suggests that tuning the parameters controlling exploration versus exploitation isn't just about optimizing engagement metrics on a test set; you have to consider how those settings impact the long-term stability and social dynamics of the user base.

Tom: Exactly. The paper is saying that relying only on one metric like watch rate isn't enough if you want a system that doesn't end up creating an extremely polarized user base.

Jane: So, in simple terms, this paper says that the way we recommend stuff matters not just for how many people click on it right now, but for how those recommendations shape what everyone thinks about in the future.

Lu: And they highlight that viral content plays a key role in this feedback loop, which is something we have to keep paying attention to as we build these systems.

Conclusion: Tom: So, we’re wrapping up on this paper about modeling how social media recommendation systems and user opinions actually feed each other in a closed loop.

Jane: It boils down to showing that the way an algorithm suggests content doesn't just change what a user sees right now, it shapes their long-term views on the whole topic.

Tom: The authors are looking at this dynamic system where users have opinions, and the recommender system keeps picking things based on those opinions.

Jane: And they found that different kinds of user opinion patterns can be really sensitive to how much influence the recommendation system has. Specifically, they looked at how polarization develops in these systems.

Tom: They tested different settings for the recommendation algorithm, like how exploratory or exploitative it is, and they saw that some setups make people's views more extreme than others.

Jane: It seems like creating viral content is a powerful way to fight off that extreme polarization they found in their models. It’s a mechanism within the system itself that can pull things back toward the middle ground.

Tom: The authors suggest there’s a sweet spot for how diverse the content recommendations are, something between too much exploration and too much exploitation, where opinions stay relatively stable.

Jane: It means that designing these systems isn't just about maximizing engagement numbers; you have to consider the social ripple effects on the users themselves.

Tom: The authors also pointed out that their watch rate modeling has a limitation—it tends to keep the user’s watch rate around fifty percent if they get content that’s neutral or close to neutral.

Jane: That's a practical piece of feedback because it tells us where our current measurement tools might be missing some nuance in how users actually interact with the stuff we recommend.

Tom: So, this research gives us a framework to test not just if an algorithm works technically, but if it creates a healthier social environment for the people using it.

Jane: And that leads us right into looking at what specific settings on those recommendation algorithms cause these polarization effects to spike or drop.

Ella C. Davidson, Mengbin Ye

Adelaide Data Science Centre · University of Adelaide

cs.SI

Submitted: 2025-07-26

Updated: 2025-11-25

Comments: 16 pages, 16 figures

Journal ref: IEEE Transactions on Computational Social Systems, 13(5): 5778-5790, July, 2026

DOI: 10.1109/TCSS.2026.3715162

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 72/100

The gist: Modelling the Closed Loop Dynamics Between a Social Media Recommender System and Users’ Opinions proposes a mathematical model to study the coupled dynamics of a Recommender System (RS) algorithm…

Key concepts

Recommender System (RS)
The RS algorithm suggests specific pieces of content to users based on their past behavior and the current social environment. It uses a payoff function to decide which piece of content is most likely to be recommended next, influencing user opinions.
User Opinion Dynamics
This describes how an individual user's belief (opinion) changes over time when they are exposed to content recommended by the system. This change is governed by how much they engage with the content and their current opinion difference from others.
Polarization and Radicalization
These terms describe a state where users' opinions become extremely divided or extreme, moving away from a neutral middle ground. The model investigates how recommender system settings affect whether the user population becomes more divided or more extreme in their views.
Viral Content
This refers to content that spreads rapidly through the social network, gaining many likes and engagement. The model shows that viral content plays a key role in stabilizing opinions by influencing users toward a consensus.

Terminology

Summary

Modelling the Closed Loop Dynamics Between a Social Media Recommender System and Users’ Opinions proposes a mathematical model to study the coupled dynamics of a Recommender System (RS) algorithm and content consumers (users) This paper finds that different opinion distributions are more susceptible to becoming polarised than others, many content stances are ineffective in changing user opinions, and creating viral content is an effective measure in combating polarisation of opinions

Model Structure

The proposed model describes a socio-technological closed-loop system where a Recommender System (RS) and users mutually influence each other The system consists of three key components: the graph representing the social media environment, the user behaviour and opinion dynamics, and the algorithm governing the RS Users hold an opinion xi(t) ∈ [−1, 1], while content has a fixed stance xcj = 2(n − 1)(j − 1) - 1 (1) The RS generates directed edges (cj, ui) from a single selected piece of content j to user i at each timestep t, forming a time-varying bipartite graph G(t)

User Behaviour Dynamics

Three key processes are modeled for a user when presented with content: the user’s engagement decision, the user’s watch rate, and the user’s opinion dynamics The engagement decision is determined by payoff functions πi(y j i(t)) dependent on the difference x¯i(t) = xi(t)−xzi(t) The probability of choosing an engagement decision y j i(t) is given by P[y j i(t) = s] = exp⁡ to βiπi(s), which is a softmax function The user watch rate wizi(t)(t) is determined by v(t) = xi(t)xzi(t), with wizi(t)(t) = 1 / (1 + e−γv(t)) The opinion evolution follows the classical Friedkin–Johnsen model, where xi(t + 1) = λi h wizi(t)(t)xzi(t) + 1 − wizi(t)(t) (5)

RS Mechanism and Payoff

The RS uses user-specific variables A j i(t), B j i(t), and C j(t) to calculate a payoff P j i(t) which determines the recommended content The payoff is defined as P j i(t) = (1 − ω)(A j i(t) + B j i(t)) squared + ωC j(t) (9) The RS then uses the softmax function to determine the probability of recommending content c j at timestep t+1: P[zi(t + 1) = cj] = exp⁡ to αP j i(t)/Pk exp⁡ to αP j i(t) (10)

Simulation Results and Findings

The Monte Carlo simulations investigate the impact of various parameters, specifically the softmax parameter α, content diversity k, and virality filtering parameters ω and δ When investigating α, smaller values correspond to an RS that aims to be more explorative (recommending a range of content) and larger values correspond to an RS that is more exploitative For the NDIC, the population becomes more polarised and radicalised as α increases, with substantial change beyond α = 16 When k = 2, the RS performs best in maximising engagement and watch rate but leads to a hyper-polarised and radicalised population For 3 ≤ k ≤ 11, the RS reduces polarisation and radicalisation in the population as opinions coalesce around the viral content When ω > 0.5, polarization and radicalisation are reduced substantially because all users are more likely to be recommended the same content

Conclusion and Limitations

Viral content emerged as a consistent phenomenon in the proposed model, its effects highlight the closed loop nature of the system The findings suggest there is a “sweet spot” of moderate content diversity (moderate k) that decreases polarisation and radicalisation A limitation of the model is that our current watch rate function results in a user always having a watch rate of approximately 50% when recommended neutral or close-to-neutral content Another limitation is that our virality-based filtering mechanism considers engagement from all users, and extending the framework to consider multiple topics could be useful for studying the rabbit-hole/spillover effect

Supplementary Material Results

The supplementary material provides results for homogenous population of λi = 0.2, showing that both NDIC and BDIC are consistently below zero, meaning polarisation and radicalisation were always decreased within the populations across the different α values For k > 9 and k is odd, there are notable decreases in the opinion distribution metrics due to a neutral piece of content going viral The results for BDIC show that the RS impacts an initially polarised population less as each user will likely already favour a particular content stance For k ≥ 10, diversity of content does not have a great effect on the BDIC and that distribution remains polarised The results for NDIC show that when k ∈ [11, 101] and the change in metrics has asymptotically diminished

Extended Results

In the extended results for Research Question Three, the contour plots exploring the relationship between ω, δ and when content goes viral within simulations for a BDIC do not substantially differ from those seen in Fig. 8 There are no notable differences between Fig. p. 2

Where the paper says it

MODEL In this section, we present and discuss the proposed model, a schematic of which is shown in Fig. There are three key components, which are covered sequentially: the graph representing the social media environment, the user behaviour and opinion dynamics, and the algorithm governing the RS.

16 and Fig. p. 2
Where the paper says it

MODEL In this section, we present and discuss the proposed model, a schematic of which is shown in Fig. There are three key components, which are covered sequentially: the graph representing the social media environment, the user behaviour and opinion dynamics, and the algorithm governing the RS.

p. 2
Where the paper says it

MODEL In this section, we present and discuss the proposed model, a schematic of which is shown in Fig. There are three key components, which are covered sequentially: the graph representing the social media environment, the user behaviour and opinion dynamics, and the algorithm governing the RS.

8 The results seen in Fig. p. 6
Where the paper says it

With α = 14, a single piece of content goes viral. As seen in Fig.

11 do not substantially differ from those seen in Fig. p. 6
Where the paper says it

With α = 14, a single piece of content goes viral. As seen in Fig.

8

Extended Results for Q3

In the extended results for Research Question Three, the contour plots exploring the relationship between ω, δ and when content goes viral within simulations for a BDIC do not substantially differ from those seen in Fig. 8 Here, ω is varied from 0 to 1, δ is varied from 1 to 1000 (note the log scale), and the colour bar refers to the proportion of the amount of likes the most liked piece of content received out of the total amount of likes recorded in the system for that simulation There are no notable differences between Fig. p. 5

Where the paper says it

To measure the performance of the RS, we record the user-specific processes of liking and watching recommended content. In each simulation, we record the total amount of likes received as a percentage of total amount of likes possible, i.

16 and Fig.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Robust recommendation strategy adjustment based on opinion distribution: The AI system can adapt its recommendation strategy by adjusting its softmax parameter alpha based on whether it is targeting a neutral or polarized population to maximize desired outcomes, as shown by as α increases, user engagement also increases for RS performance metrics.

  2. Content diversity optimization for polarization control: The system can dynamically adjust the number of content stances k to find a sweet spot that decreases polarization and radicalization in users' opinions while maintaining high engagement, specifically noting that "when 3 <= k <= 11, the RS, in fact, reduces polarisation and radicalisation in the population."

  3. Virality-aware recommendation weighting: The AI can integrate virality into its decision-making by adjusting the weight omega to balance content-based filtering with viral content influence; increasing omega above 0.5 is shown to reduce polarization and radicalization in a way that all users are more likely to be recommended the same content.

  4. Dynamic memory management for viral detection: The system can adjust the memory size delta when virality is being used, as increasing delta beyond 1 leads to a sharp decrease in the level of virality, indicating a mechanism to prevent over-reliance on short-term viral spikes.

  5. Personalized engagement modeling: The AI can model user engagement decisions using payoff functions based on opinion difference, allowing it to predict whether a user will like, dislike, or remain neutral content based on the current recommendation and their existing opinion.

Abstract

This paper proposes a mathematical model to study the coupled dynamics of a Recommender System (RS) algorithm and content consumers (users). The model posits that a large population of users, each with an opinion, consumes personalised content recommended by the RS. The RS can select from a range of content to recommend, based on users' past engagement, while users can engage with the content (like, watch), and in doing so, users' opinions evolve. This occurs repeatedly to capture the endless content available for user consumption on social media. We employ a campaign of Monte Carlo simulations to study how recommender systems influence users' opinions, and in turn how users' opinions shape the subsequent recommended content. Both the performance of the RS (e.g., how users engage with the content) and the polarisation and radicalisation of users' opinions are of interest. We find that different opinion distributions are more susceptible to becoming polarised than others, many content stances are ineffective in changing user opinions, and creating viral content is an effective measure in combating polarisation of opinions. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

Sources

Related papers