Media Environment, Dual Process and Polarization: A Computational Approach

TL;DR

This paper models how media environments influence political polarization using a Bayesian network that combines motivated and analytic reasoning, showing polarization worsens in a post-truth context.

Contribution

It introduces a novel computational model integrating dual reasoning theories to analyze media influence on polarization.

Findings

Polarization increases in post-truth media environments

Bayesian network effectively captures media influence dynamics

Model highlights importance of truthfulness and alignment in media

Abstract

News articles of varying degrees of truthfulness and political alignment, and their influences on the political opinions of the media consumers are modeled as a Bayesian network incorporating a mixture of ideas from dual-reasoning models of Motivated Reasoning and Analytic/Intuitive Reasoning. The result shows that as the media environment moves towards the Post-Truth world, the problem of political polarization becomes exacerbated.

Figures

Equations

$$\mathbb{P}\left(p_n\mid X=1\right)$$

$$\mathbb{P}\left(p_n\mid X=0\right)$$

$$\mathbb{P}\left(X=0\right)=$$

$$d_m$$

$$x_n$$

$$x_a$$

$$\mathbb{P}\left(t_j=\text{true}\mid x_n>x_a\right)$$

$$\mathbb{P}\left(t_j=\text{true}\mid x_n\le x_a\right)$$

$$p_j$$

$$\mathbb{P}\left(p_a\mid p_j\right)$$

Full text

1 Introduction

“Fake News” came to the spotlight in a recent surge of the use of the term. It has been at least partially attributed to for recent upsets in elections (Rose (\APACyear2017)), Anti-Vaxxers and other movements based on false medical information (Waszak \BOthers. (\APACyear2018)), and global warming denial (van der Linden \BOthers. (\APACyear2017)). Fake News is also blamed for the radicalization of fringe movements and the rise of racially and religiously motivated hate-crimes (Townsend (\APACyear2017); Hermansson (\APACyear2018)).

In this introductory section, we first will set the scene by characterizing the recent development of Fake News and the new media environment. Then, we will cite studies that are of particular interest to the present study. Finally, we will give an overview of the plan to model and measure the effect of the media.

1.1 The new media environment

Two features of the new development are of interest to us. First is the relativization of the concept of Truth. In the post-truth politics, the line between assertions and emotional fleas and the truth is blurred. With the advent of the new media environment, traditional journalists are no longer seen as the gatekeepers of the fact-based public opinion, and the notion of Truth has lost its privileged in the media landscape, and we have seen an increase in media outlets that capitalizes on various propaganda of questionable truthfulness. (Suiter (\APACyear2016), Gorrell \BOthers. (\APACyear2019)).

Second is the polarization effect. It has long been observed that the advent of the new media environment has the effect of accelerating the polarization of public opinion (Spohr (\APACyear2017)). From the “One-Step flow” point of view that the news readers exposure to the media directly shape the news reader’s own opinion (Bennett \BBA Manheim (\APACyear2006)), as the political alignment of post-truth media approaches fringe politics, readers of such news media become more polarized. On top of that, various back-firing effects have been observed, whereby readers of particular political mindset exercise Motivated Reasoning in discarding the opposite side of the story and further reinforces the existing belief.

These two characteristics of the new media have pushed the political landscape to an alarming level of radicalization, as observed from various election upsets to the rise of hate crime and other consequences of the fringe movements. In this paper, we will model the different media environments by allocating media consumption to premium centrist media, premium partisan media and Fake News media. We will discuss the media environment model in detail in the Methods section.

1.2 Recent studies on political information consumption

We list recent studies to draw inspiration from, to model the behavior of the media consumers.

Taber \BOthers. (\APACyear2009) show the effect of dis-confirmation bias of Motivated Reasoning. The result suggests that, when processing political news, readers tend not to exercise the same level of scrutiny to determine the truthfulness of the news item.

Pennycook \BBA Rand (\APACyear2018) found that the ability to tell factually accurate news from Fake News is positively correlated to the participants’ performance in the Cognitive Reflection Test. This result suggests that Motivated Reasoning alone is not sufficient to account for the discrepancies between individual media consumer’s ability to single out Fake News.

Rather than completely replacing the existing observations on Motivated Reasoning, we will use the result as a complementary force in determining the probabilistic degree to which the agent will allow factual inaccuracies, as we shall see in the following section.

Bail \BOthers. (\APACyear2018) found evidence for the backfiring effect, whereby participants who were regularly exposed to the opposing views were found to be even more polarized. In this study, to account for such backfiring, we will count those that the agent will deem as Fake News to have the opposite effect. This Fake News backfiring effect, combined with the Motivated Reasoning phenomena, should be able to account for the backfiring effect in general. We won’t introduce a sub-module for backfiring in general, however, as we believe that Motivated Reasoning

• Fake News Backfiring should suffice to account for the backfiring results in general, given the current media environment.

Now that we have examined a couple of recent empirical studies, we will turn to the theoretical underpinning of the present study.

1.3 Polarization and Bayesian network

Jern \BOthers. (\APACyear2014) found that Bayesian networks that have root nodes other than the hypothesis node can capture the polarization or convergence effect, where two agents of diverging effects will either be polarized or converge to the center after examining the same data point. As we shall see later in this paper, the topology of the Bayesian network presented in this paper has the “truth judgment” intervening node that encompasses Motivated Reasoning and Fake News Backfiring. Unlike Jern \BOthers. (\APACyear2014), however, we use explicit programming constructs available to us in a probabilistic programming language to model the backfiring effect, and we do so on a continuous distribution. Nevertheless, the basic idea that the intervening root node in a Bayesian network can account for polarization remains intact.

2 Methods

In this section, we explain the design of the model and give the rationale for design decisions. On the media side, we will first examine the definition of a news item. Then we will discuss the abstraction of a media outlet, and we will finally discuss the media environment. On the agent side, we will discuss a Bayesian network in which the truthfulness of a news item is determined by an agent.

2.1 News item

A news item has two numerical properties: politics $\left(p_{n}\right)$ and truth $\left(t_{n}\right)$.

$p_{n}$ value is centered at 0. The further away from the center the politics value is, the more extreme the political alignment of the news item is. If the absolute value is close to 1 or above, the politics of the news item is considered pretty extreme.

$t_{n}$ value is 0 or above. Truth value of 1 or above is considered factually accurate in every detail. Being true to the concept of Post-Truth politics, $n_{t}$ is a continuous variable, rather than a Boolean value.

2.2 Media outlet

Each type of media outlet emits a news item of probabilistic politics and truth values. We define three different types of media outlets and their $p_{n}$ and $t_{n}$ distributions.

[TABLE]

For partisan media outlets, their $p_{n}$ is bi-modal with equal weight on both sides i.e. for Premium Partisan media:

[TABLE]

Similarly, Fake News Partisan media is also bi-modal, with more extreme means.

2.3 Media environment

Media environment defines a mixture of media outlets. At each observation, the media environment makes a weighted random choice among its mixture of outlet, and let the chosen outlet to generate a news item. Here are the media environments and their mixtures:

[TABLE]

ME1 represents an ideal media environment, with a majority of centrist media, with a few partisan media outlets and very few Fake News media. ME2 represents an environment where media outlets are more opinionated. ME3 represents a post-truth media environment with an alarmingly large amount of Fake News circulation. Such shift of media environments from ME1 to ME3 over time can be attributed to the economic reality for news media to find an audience in the increasingly crowded media industry (Bernhardt \BOthers. (\APACyear2008); Posner (\APACyear2005)).

Now we turn to explain how agent processes news items.

2.4 Agent model

An agent has two numerical properties: politics $\left(p_{a}\right)$ and analytic $\left(a_{a}\right)$. Following Pennycook \BBA Rand (\APACyear2018), we assign the $a_{a}$ value as an intrinsic property of an agent. While $p_{a}$ gets updates as the agent processes news items, the analytic value stays the same throughout the process.

Like news item, the prior distribution of $p_{a}$ of an agent is centered around 0. The prior distribution is $p_{a}\sim\mathcal{N}\left(0,1\right)$. $a_{a}$ value is sampled from $a_{a}\sim\text{Uniform}\left(0.5,1\right)$.

2.5 Truth and politics judgment by agent

Upon observing a news item, the agent makes a Boolean decision of whether the news item is true or false. In order to make the decision, the agent first calculates the motivational discount $d_{m}$:

[TABLE]

That is, the discount exponentially decays as the $p_{n}$ gets further away from $p_{a}$. This reflects the observations of Motivated Reasoning that agents are less likely to call out Fake News that are more aligned to their own politics.

Two variables are sampled:

[TABLE]

truth judgment $\left(t_{j}\in\left\{\text{true},\text{false}\right\}\right)$ is true iff $x_{n}>x_{a}$:

[TABLE]

Politics judgment $\left(p_{j}\right)$ reflects the agent’s perception of the political effect of the news item. When $t_{j}$ is false, $p_{j}$ is the reverse of $p_{n}$.

[TABLE]

This formulation reflects the backfiring effect of perceived Fake News.

2.6 Updating $p_{a}$

$p_{j}$ is the observation value with which $p_{a}$ is updated. The formulation of the Bayesian update is as following:

[TABLE]

where the likelihood function $\mathbb{P}\left(p_{j}\mid p_{a}\right)$ is the pdf of distribution $P\sim\mathcal{N}\left(p_{a},0.25\right)$

In summary, Figure 1 shows the model as a Bayesian network.

2.7 Implementation of the model

We used Anglican probabilistic programming language (van de Meent \BOthers. (\APACyear2018)) to program the model. In order to obtain the resulting distribution, we have run the model with Parallel Lighweight Metropolis-Hastings algorithm with 5,000 particles, and obtained the first 200,000 iterations and drew the histograms of the results111Source code for the model is available at https://gitlab.com/chajath/polarize-fakenews-model.

3 Results

Figures 2 shows $P_{a}$ distributions after 1, 10 and 100 observations of news item for given media environments.

ME1 shows a Gaussian distribution of $p_{a}$ after prolonged exposure to the environment. The strong majority of opinions are formed within a moderate range of the political spectrum (i.e. within $\pm 0.5$) This is hardly a surprising result given the majority of the media outlet in ME1 falls under such political spectrum.

ME2 is a more interesting case. Early distributions after 1 and 10 observations show a clear bi-modality of the distribution, which is not surprising given the large portion of the partisan media. After 100 observations, however, the distribution looks a lot like ME1 distribution, with a strong majority opinion formed within a moderate range. From this result, it stands to reason that media taking editorial stances on issues are healthy phenomena given that they are factual, and the overall distribution of the partisan media is equally distributed on both sides of the argument.

ME3 results show a high degree of polarization in the extreme range of the political spectrum. After 10 observations, an alarming level of concentration of the opinion is found on both extremes. This is different from ME1 and ME2 results where you can still find a strong majority within the moderate band. This shows the loss of common ground in the public discourse in a Post-Truth politics. After 100 observations, a moderate opinion does gain some of the shares back. However, the majority is opinions are still outside of the moderate band, with a very high concentration found on each of the extreme sides.

Overall, the result shows a detrimental effect that Fake News has on public discourse according to our model. The result can account for an unusually high level of opinion concentration on fringe opinion, and an upsetting result of the public poll when the proposal is a rather extreme one.

4 Discussion and Conclusion

The biggest limitation of this study is the lack of any data fitting in designing the model. Although we did consider empirical results from other studies for inspiration, the model itself is largely hand-wired with our subjective view and understanding of the matter. In order to tune our model to reflect the reality better, many of the magic numbers appearing in our model need to be parameterized: {APAitemize}

prior distribution of $p_{a}$ and $a_{a}$

scale and decay rate in $d_{m}$ formulation

standard deviation in likelihood function

To fit those values, in the future study, we shall gather statistics from existing surveys of the media environment, and design controlled experiment to determine the parameters within the model. Finally, another experiment shall be conducted to determine the accuracy of the model compared to real-world experimental results.

Although the prescriptive analysis was not the main theme of this study, the study can be further extended to account for such analysis. The conclusion from studies with a prescriptive focus such as van der Linden \BOthers. (\APACyear2017) and McClain (\APACyear2017) suggest that an antidote to Fake News is the effective communication of scientific consensus. Given the recurrent structure of our model, this can be translated to observing moderate news items with high truth values prior to being exposed to Fake News items.

In this study, we have essentially abstracted out news items as two variables: politics and truthfulness. In the future study, it will be interesting to give a more detailed account of the semantics of news item. Given the recurrent structure of the model, it will be natural to extend the model to interface with other types of Bayesian recurrent neural network (Fortunato \BOthers. (\APACyear2017)) that focuses on natural language processing. Doing so will allow us to train the model from primary sources of the data, namely the news items themselves.

What we have presented here is a way to model the behavior of individual news-consuming agents. Given the advent of personal media and the prevalence of social media, it will be interesting to devise a multi-agent model (following Aymanns \BOthers. (\APACyear2017)) where the interaction between the agents are modeled after social networks, with varying degrees of reaches and influences from multiple individuals.

The present study showcases an interesting use of probabilistic programming to model complex social phenomena from the cognitive perspective of individual agents. Future work ideas discussed in this section will further the use of probabilistic programming and other computational cognitive methods in the realms of computational social sciences.