Operational Definitions.

To explore the key differences between social media platforms and how they influence echo chambers’ formation, we need to operationalize a definition for them. First, we need to identify the attitude of users at a microlevel. On online social media, the individual leaning of a user $$ i$$ toward a specific topic, $$ x_i$$, can be inferred in different ways, via the content produced or the endorsement network among users (37). Concerning content, we can define the leaning as the attitude expressed by a piece of content toward a specific topic. This leaning can be explicit (e.g., arguments supporting a narrative) or implicit (e.g., framing and agenda setting). Let us consider a user $$ i$$ producing a number $$ a_i$$ of contents, $$ {\mathcal{C}}_i=\left\{c_1,c_2,\dots ,c_{a_i}\right\}$$, where $$ a_i$$ is the activity of user $$ i$$, and each content leaning is assigned a numeric value. Then the individual leaning of user $$ i$$ can be defined as the average of the leanings of produced contents,

Implementation on Social Media.

This section explains how we implement the operational definitions defined above on different social media. For each medium, we detail 1) how we quantify users’ leaning, and 2) how we reconstruct how the information spread.

Polarization and Homophily in the Interaction Networks.

The network’s topology can reveal echo chambers, where users are surrounded by peers with similar leanings, and thus they get exposed, with a higher probability, to similar contents. In network terms, this translates into a node $$ i$$ with a given leaning $$ x_i$$ more likely to be connected with nodes with a leaning close to $$ x_i$$ (18). This concept can be quantified by defining, for each user $$ i$$, the average leaning of their neighborhood, as $$ x_i^N\equiv \frac{1}{k_i^{\to }}{\sum }_j{A}_{ij}{x}_j$$, where $$ A_{ij}$$ is the adjacency matrix of the interaction network, $$ A_{ij}=1$$ if there is a link from node $$ i$$ to node $$ j$$, $$ A_{ij}=0$$ otherwise, and $$ k_i^{\to }={\sum }_j{A}_{ij}$$ is the out-degree of node $$ i$$. Fig. 1 shows the correlation between the leaning of a user $$ i$$ and the leaning of their neighbors, $$ x_i^N$$, for the four social media under consideration. The probability distributions $$ P\left(x\right)$$ (individual leaning) and $$ P^N\left(x\right)$$ (average leaning of neighbors) are plotted on the $$ x$$ and $$ y$$ axes, respectively. All plots are color-coded contour maps, representing the number of users in the phase space $$ \left(x,x^N\right)$$: The brighter the area in the plan, the larger the density of users in that area. The topics of vaccines and abortion, on Facebook and Twitter, respectively, show a strong correlation between the leaning of a user and the average leaning of their nearest neighbors. Similar behavior is found for different topics from the same social media platform; see SI Appendix. Conversely, Reddit and Gab show a different picture. The corresponding plots in Fig. 1 display a single bright area, indicating that users do not split into groups with opposite leaning but form a single community, biased to the left (Reddit) or the right (Gab). Similar results are found for different datasets on Reddit; see SI Appendix.

Fig. 1.

Joint distribution of the leaning of users $$ x$$ and the average leaning of their neighborhood $$ x^{NN}$$ for different datasets. (A) Twitter, (B) Reddit, (C) Facebook, and (D) Gab. Colors represent the density of users: The lighter the color, the larger the number of users. Marginal distribution $$ P\left(x\right)$$ and $$ P^N\left(x\right)$$ are plotted on the x and y axes, respectively. Facebook and Twitter present by homophilic clustering.

The presence of homophilic interactions can be confirmed by the community structure of the interaction networks. We detected communities by applying the Louvain algorithm (41), removing singleton communities with only one user. Then, we computed each community’s average leaning, determined as the average of individual leanings of its members. Fig. 2 shows the communities emerging for each social medium, arranged by increasing average leaning on the $$ x$$ axis (color-coded from blue to red), while the $$ y$$ axis reports the size of the community. On Facebook and Twitter, communities span the whole spectrum of possible leanings, but users with similar leanings form each community. Some communities are characterized by a robust average leaning, especially in the case of Facebook. These results are in accordance with the observation of homophilic interactions. Instead, communities on Reddit and Gab do not cover the whole spectrum, and all show similar average leaning. Furthermore, the almost total absence of communities with leaning very close to 0 confirms the polarized state of the systems.

Fig. 2.

Size and average leaning of communities detected in different datasets. A and C show the full spectrum of leanings related to the topics of abortions and vaccines with regard to communities in B and D, where the political leaning is less sparse.

Effects on Information Spreading.

Simple models of information spreading can gauge the presence of echo chambers: Users are expected to be more likely to exchange information with peers sharing a similar leaning (18, 42, 43). Classical epidemic models such as the susceptible–infected–recovered (SIR) model (44) have been used to study the diffusion of information, such as rumors or news (4546–47). In the SIR model, each agent can be in any of three states: susceptible (unaware of the circulating information), infectious (aware and willing to spread it further), or recovered (knowledgeable but not ready to transmit it anymore). Susceptible (unaware) users may become infectious (aware) upon contact with infected neighbors, with a specific transmission probability $$ \beta $$. Infectious users can spontaneously become recovered with probability $$ \nu $$. To measure the effects of the leaning of users on the diffusion of information, we run the SIR dynamics on the interaction networks, by starting the epidemic process with only one node $$ i$$ infected, and stopping it when no more infectious nodes are left.

The set of nodes in a recovered state at the end of the dynamics started with user $$ i$$ as a seed of infection, that is, those that become aware of the information initially propagated by user $$ i$$, form the set of influence of user $$ i$$, $$ {\mathcal{I}}_i$$ (48). Thus, the set of influence of a user represents those individuals that can be reached by a piece of content sent by him/her, depending on the effective infection ratio $$ \beta /\nu $$. One can compute the average leaning of the set of influence of user $$ i$$, $$ {\mu }_i$$, as

Fig. 3 shows the average leaning $$ ⟨\mu \left(x\right)⟩$$ of the influence sets reached by users with leaning $$ x$$, for the different datasets under consideration. The recovery rate $$ \nu $$ is fixed at 0.2 for every dataset. In contrast, the ratio between the infection rate $$ \beta $$ and average degree $$ ⟨k⟩$$ depends on the specific dataset and is reported in the caption of each figure.

Fig. 3.

Average leaning $$ ⟨\mu \left(x\right)⟩$$ of the influence sets reached by users with leaning $$ x$$, for different datasets under consideration. Size and color of each point represent the average size of the influence sets. The parameters of the SIR dynamics are set to (A) $$ \beta =0.10{⟨k⟩}^{-1}$$, (B) $$ \beta =0.01{⟨k⟩}^{-1}$$, (C) $$ \beta =0.05{⟨k⟩}^{-1}$$, and (D) $$ \beta =0.05{⟨k⟩}^{-1}$$, while $$ \nu $$ is fixed at 0.2 for all simulations.

Again, one can observe a clear distinction between Facebook and Twitter, on one side, and Reddit and Gab on the other side. For the topics of vaccines and abortion, on Facebook and Twitter, respectively, users with a given leaning are much more likely to be reached by information propagated by users with similar leaning, that is, $$ ⟨\mu \left(x\right)⟩\approx x$$. Similar behavior is found for different topics from the same social media platform; see SI Appendix. Conversely, Reddit and Gab show a different behavior: The average leaning of the set of influence, $$ ⟨\mu \left(x\right)⟩$$, does not depend on the leaning $$ x$$. As expected, the average leaning in these media is not zero. Still, it assumes negative (positive) values in Reddit (Gab), indicating that the users of this platform are more likely to receive left (right)-leaning content.

These results indicate that information diffusion is biased toward individuals who share a similar leaning in some social media, namely Twitter and Facebook. In contrast, in others—Reddit and Gab in our analysis—this effect is absent. Such a latter configuration may depend upon two factors: 1) Gab and Reddit are not bursting the echo chamber effects, or 2) we are observing the dynamic inside a single echo chamber.

Our results are robust for different values of the effective infection ratio $$ \beta /\nu $$; see SI Appendix. Furthermore, Fig. 3 shows that the spreading capacity, represented by the average size of the influence sets (color-coded in Fig. 3), depends on the leaning of the users. On Twitter, proabortion users are more likely to reach larger audiences. The same is true for antivax users on Facebook, left-leaning users on Reddit, and right-leaning users on Gab (in this dataset, left-leaning users are almost absent).

News Consumption on Facebook and Reddit.

The striking differences observed across social media, in terms of homophily in the interaction networks and information diffusion, could be attributed to the different topics taken into account. For this reason, here we compare Facebook and Reddit on a common topic, news consumption. Facebook and Reddit are particularly apt to a cross-comparison since they share the definition of individual leaning (computed by using the classification provided by mediabiasfactcheck.org; see Materials and Methods for further details) and the rationale in creating connections among users that is based on an interaction network. Fig. 4 shows a direct comparison of news consumption on Facebook and Reddit along the metrics used in the previous sections to quantify the presence of echo chambers: 1) the correlation between the leaning of a user $$ x$$ and the average leaning of neighbors $$ x^N$$ (Fig. 4, Top), 2) the average leaning of communities detected in the networks (Fig. 4, Middle), and 3) the average leaning $$ ⟨\mu \left(x\right)⟩$$ of the influence sets reached by users with leaning $$ x$$, by running SIR dynamics (Fig. 4, Bottom). One can see that all three measures confirm the picture obtained for other datasets: On Facebook, we observe a clear separation among users depending on their leaning, while, on Reddit, users’ leanings are more homogeneous and show only one peak. In the latter social media, even users displaying a more extreme leaning (noticeable in the marginal histogram of Fig. 4 B, Top) tend to interact with the majority. Moreover, on Facebook, the seed user’s leaning affects who the final recipients of the information are, therefore indicating the presence of echo chambers. On Reddit, this effect is absent.

Fig. 4.

Direct comparison of news consumption on (A) Facebook and (B) Reddit. Joint distribution of the leaning of users $$ x$$ and the average leaning of their nearest neighbor $$ x^N$$ (Top), size and average leaning of communities detected in the interaction networks (Middle), and average leaning $$ ⟨\mu \left(x\right)⟩$$ of the influence sets reached by users with leaning $$ x$$, by running SIR dynamics (Bottom) with parameters $$ \beta =0.05⟨k⟩$$ for A, $$ \beta =0.006⟨k⟩$$ for B, and $$ \nu =0.2$$ for both. Facebook presents a highly segregated structure with regard to Reddit.

Labeling of Media Sources.

The labeling of news outlets is based on the information reported by Media Bias/Fact Check (MBFC) (https://mediabiasfactcheck.com), an independent fact-checking organization that rates news outlets on the basis of the reliability and of the political bias of the contents they produce and share. The labeling provided by MBFC, retrieved in June 2019, ranges from Extreme Left to Extreme Right for political bias. The total number of media outlets for which we have a political label is 2,190. A detailed description of the source labeling process and political bias distribution can be found in SI Appendix.

Data Availability.

For what concerns Gab, all data are available on the Pushshift public repository (https://pushshift.io/what-is-pushshift-io/) at this link: https://files.pushshift.io/gab/. Reddit data are available on the Pushshift public repository at this link: https://search.pushshift.io/reddit/. For what concerns Facebook and Twitter, we provide data according to their Terms of Services on the corresponding author institutional page at this link: https://walterquattrociocchi.site.uniroma1.it/ricerca. For news outlet classification, we used data from MBFC (https://mediabiasfactcheck.com), an independent fact-checking organization. Anonymized data have been deposited in Open Science Framework (10.17605/OSF.IO/X92BR) (49). For further details about data, refer to the following section.

Empirical Datasets.

Table 1 reports summary statistics of the datasets under consideration. Due to the structural differences among platforms, each dataset has different features. For Twitter, we used tweets regarding three topics collected by Garimella et al. (16), namely Gun control, Obamacare, and abortion. Tweets linking to a news source with a known bias are classified based on MBFC. Facebook datasets were created by using Facebook Graph API and were previously explored in ref. 50 (Science and Conspiracy), (51) (Vaccines) and (11) (News). For the two datasets Science and Conspiracy and Vaccines, data were labeled in a binary way, respectively, provaccines/antivaccines and proscience/conspiracy, based on the page where they were posted. Posts in the dataset News were instead classified based on MBFC labeling. Reddit datasets have been obtained by downloading comments and submissions posted in the subreddit Politics, the_donald, and News and labeled according to the classification obtained from MBFC. The Gab dataset has been collected from https://files.pushshift.io/gab and contains posts, replies, and quotations. Posts were labeled according to MBFC classification. Further details can be found in SI Appendix.

Table 1.

Dataset details

Media	Dataset	$$ T_0$$	$$ T$$	$$ C$$	$$ N$$	$$ n_c$$
Twitter	Gun control	June 2016	14 d	19 million	3,963	0.93
	Obamacare	June 2016	7 d	39 million	8,703	0.90
	Abortion	June 2016	7 d	34 million	7,401	0.95
Facebook	Sci/Cons	January 2010	5 y	75,172	183,378	1.00
	Vaccines	January 2010	7 y	94,776	221,758	1.00
	News	January 2010	6 y	15,540	38,663	1.00
Reddit	Politics	January 2017	1 y	353,864	240,455	0.15
	the_donald	January 2017	1 y	1.234 million	138,617	0.16
	News	January 2017	1 y	723,235	179,549	0.20
Gab	Gab	November 2017	1 y	13 million	165,162	0.13

For each dataset, we report the starting date of collection $$ T_0$$, time span $$ T$$ expressed in days (d) or years (y), number of unique contents $$ C$$, number of users $$ N$$, and coverage $$ n_c$$ (fraction of users with classified leaning). For Twitter, $$ T$$ represents the window to sample active users, of which we retrieve all of the tweets related to the topic via the Application Programming Interface (API) (more information in SI Appendix). Sci/Cons, Scientific and Conspiracy content.