Modeling Diffusion of Information in an Increasingly Complex Digital Domain

Offering entertainment, discussion, and information, social media provides users with a stimulating online experience. Within the last five years, it has also become an increasingly popular medium for the consumption of news. News outlets publish articles and reports through social media, and by doing so influence their users in a way that corresponds with the outlet’s political leaning. Because social media outlets provide users with tailored content, the prevalence of biased news reporting reinforces the user’s political values and polarizes their beliefs. This thesis attempts to examine the relationships that give rise to this political polarization in social media and discusses possible opportunities to mitigate it.


Introduction
This paper will use the system dynamics modeling process as outlined by John Sterman of MIT: articulating the problem, formulating a dynamic hypothesis, formulating a simulation model, testing that model, formulating effective policies, and evaluating those policies (Sterman, 2000).

Problem Articulation
In today's world, personalization has become a cornerstone of the online experience.Social media giants, in particular, create individualized experiences for their users based on their expressed preferences, historical trends, and predicted interests (Van Dijck & Poell, 2013).When applied to politics, these environments inundate users with similar or related information that reinforces their current political beliefs.This inundation isolates users from alternative opinions and viewpoints, polarizing their perspective.This polarizing isolation effect is commonly referred to as the "echo chamber" or "filter bubble" (Karsten & West, 2016).The challenge addressed in this paper to illuminate the relationships that give rise to this political polarization in social media and to discuss possible opportunities to mitigate it.

Variables
The variables within the echo effect model include:

Testing
The two variables manipulated were the Random Number Generator (RNG) Seed and Media News Reporting.These two variables fed values into the other functions within the model, impacting on the resulting behavior.The first variable addressed was the RNG Seed.The RNG Seed variable feeds into the Random Draw function, which dictates the effectiveness of the personalization algorithms.The RNG Seed has a range between 0 and 50 and must be a whole number.For the extreme condition testing, our team varied the RNG Seed value within this range and compared it to a baseline middle value of 25.Below is a graph of the resulting behavior our team observed in the proportion of Right-Wing Beliefs.As can be seen in Figure 4, despite the RNG Seed value used, a linear trend of Left-Wing Articles Consumed increasing over time still exists.This indicates that the RNG Seed number has no significant effect on the long-term behavior of Left-Wing Articles Consumed.This behavior is similar in Right-Wing Articles Consumed as well, however, the number of Right-Wing Articles Consumed is far fewer than the number of Left-Wing Articles Consumed.
The Media News Reporting variable feeds both the Flow of Left-Wing Tailored Content and Flow of Right-Wing Tailored Content, controlling the number of articles a user internalizes and is measured as Articles/Day.Preliminary testing of this variable showed that the approximate range for this variable was between 0 and 15, as increases above 15 negligibly affected the speed to which a user becomes polarized.So for extreme conditions testing, Media News Reporting was varied between 0 and 15. Figure 5

Sensitivity Analysis
After extreme conditions testing, sensitivity analysis was conducted on the same variables outlined in the extreme conditions testing, but within a smaller range to observe changes in behavior.The first variable tested was the RNG Seed.The RNG Seed variable feeds into the Random Draw function, which dictates the effectiveness of the personalization algorithms.The RNG Seed has a range between 0 and 50 and must be a whole number.For the sensitivity analysis, our team varied the RNG Seed value within the confines of this range.We chose to use RNG Seed values of 0, 10, 20, 30, 40, and 50.Below is a graph of the resulting behavior our team observed in the proportion of Right-Wing Beliefs.As can be seen in the graph, there is no clear relationship between the RNG Seed value and the resulting behavior of the Proportion of Right-Wing Beliefs.This is consistent with our findings outlined in the extreme conditions testing section.This similar behavior can be observed in the complementary graph for Proportion of Left-Wing Beliefs, which is depicted below.The next indicator analyzed was Articles Consumed.

Policy Design and Evaluation
The first proposed policy recommendation is to institute a federal restriction capping the effectiveness of personalization algorithms.This artificially imposed cap would force social media giants to either allow or feed media news reporting of the opposite bias to their users.
The second proposed policy recommendation mandates that social media giants include a time delay within their personalization algorithms.This time delay would ultimately delay the polarization process, allowing users to experience more varied media news reporting in the newly available time through their own exploration.This additional time for extra exposure would cause the personalization algorithms to introduce less polarized material to their users.

Conclusion
The model demonstrates that polarization does occur.This follows the principles of the Polya process and path dependence.Therefore, it passes a general common sense test for expected model behavior.This modeled polarization is important for consumers to understand because it not only affects the information they receive, but also has the potential to alter their political opinions and thereby their decisions and actions.Consumers of social media should diversify their intake of information to reduce the effect of polarization and promote a political alignment that is uninfluenced by social media giants.Both proposed policy recommendations have the potential to facilitate that intake diversification, promoting a less polarized-and hopefully more open, educated, and well-rounded population.

Figure 1 .
Figure 1.Reference Mode of Proportion of Right-Wing Beliefs

Figure
Figure 2. Stock and Flow Diagram

Figure 3 .
Figure 3. Effects of RNG Seed Extreme Conditions Testing on Proportion of Left and Right-Wing Beliefs

Figure 4 .
Figure 4. Effects of RNG Seed Extreme Conditions Testing on Left-Wing and Right-Wing Articles Consumed depicts the relationships between Media News Reporting and the Proportion of Left-Wing Belief sand the Proportion of Right-Wing Beliefs.

Figure 5 .
Figure 5. Effects of Media News Reporting Extreme Conditions Testing on the Proportion of Left-Wing Beliefs and the Proportion of Right-Wing Beliefs

Figure 6 .
Figure 6.Effects of Media News Reporting Extreme Conditions Testing on Right-Wing Articles Consumed

Figure 7 .
Figure 7. Proportion of Left-Wing Beliefs and Proportion of Right-Wing Belief RNG Seed Sensitivity

Figure 9 .
Figure 9. Proportion of Left-Wing Beliefs and Right-Wing Beliefs Media News Reporting Sensitivity

Figure 10 .
Figure 10.Left-Wing Articles Consumed and Right-Wing Articles Consumed Media News Reporting Sensitivity User Political Interests, Personalization Algorithms, Media News Reporting, Right-Wing Beliefs, Left-Wing Beliefs, Proportion of Right-Wing Beliefs, Proportion of Left-Wing Beliefs, Flow of Right-Wing Tailored Content, and Flow of Left-Wing Tailored Content.See Table1for their definitions.Media content not derived from social media is not considered in this model.
Table2depicts each variable, its description, units, associated equation, initial value, and supporting logic.