Conflicts and interactions between subreddits, how the cooperation and conflict dynamics start and evolve in time.
ada-2025-project-additionalgroup2
Social ecosystems are ruthless. Tensions emerge, alliances form, conflicts rage and never cease.
With the advent of the Internet, interconnected social ecosystems have emerged, ushering in a new era of confrontations and incessant rivalries.
Among them, Reddit stands out for its notoriety and belligerent reputation.
Behind its seemingly harmless appearance as a simple forum lies a true digital battlefield.
Although the dynamics that animate it are highly complex, Reddit is built upon a profoundly simple structure.
It is a giant forum, composed of thousands of autonomous groups called subreddits, each centered around a specific theme.
There are all kinds, and when one gains popularity, it generates a true community.
Between these different communities, links are forged, discussions emerge, mutual aid appears, and provocations are launched.
This interconnected structure, galvanized by anonymity, makes Reddit a perfect example of social dynamics where freedom is absolute.
In this organized chaos, several questions arise:
- Do these interactions obey hidden laws?
- Is Reddit a network dominated by a few centers of power, or a more egalitarian ecosystem?
- Who attacks, who suffers, and why?
- How do these conflicts evolve over time?
To explore these dynamics, we analyzed over a million links between subreddits,
each representing a mention or citation from one group to another.
These links come from a dataset produced by the Pushshift project, covering the period from 2014 to 2017.
Each link contains:
- a source subreddit
- a target subreddit
- a date
- the nature of the interaction (positive or negative)
- a vector representing the text properties of the post
After merging the different files, removing duplicates, and normalizing the formats,
we were able to analyze these interactions on a large scale.
We began our study with an analysis of the global structure of exchanges between communities.
We identified the main actors and the main aggressors.
We examined whether there was reciprocity between the number and nature of outgoing and incoming links.
We also analyzed reciprocity between the attitude of a subreddit toward others, and that of others toward it.
Finally, we looked into the distribution of conflicts:
Are they concentrated? Dispersed? Can we detect power dynamics within them?
We then continued with a thematic analysis,
seeking to determine whether the theme of a subreddit influences the nature of its interactions.
- Does thematic proximity between subreddits influence the frequency of their exchanges?
- Or the degree of conflictuality?
- More broadly, does a subreddit’s theme have a structuring effect on the network and its dynamics?
We then focused on the temporal dimension.
By comparing the years 2014 to 2017, we analyzed how interactions evolve over time:
- The effect of time on the intensity of exchanges
- The evolution of conflicts
- The stability of relationships
Do some dynamics become lasting?
Does this stability depend on the nature of interactions?
After that, we extended our perspective beyond thematic similarities,
to explore semantic similarities between subreddits.
We tried to understand:
- How they are formed
- How they interact between each other
- To what extent they help shape the structure and nature of interactions
Finally, we studied phenomena of emergence:
- What types of events or content can act as a trigger?
- What is the intensity of the propagation they generate?
- To what extent do the resulting dynamics become structuring?
Are there recurring patterns in the way these chains of interactions propagate through the network?
Every social dynamic relies on a structure, and it is through its study that we initiated our analysis.
Reddit is not a homogeneous space: some subreddits occupy more room, others are practically nonexistent.
Some are violent, while others are purely pacifist.
To what extent are these behaviours symmetrical, and are the interaction intensities balanced?
In this first axis of analysis, we sought to understand how interactions and their nature are organised across multiple scales.
By analysing the structure of the network and the distribution of exchanges, we attempted to identify the logics of power and reciprocity that govern this ecosystem.
In order to begin our analysis of the structure of Reddit’s social ecosystem, we conducted a preliminary analysis of the dataset and obtained some basic but nevertheless important information, such as the list of the most active and most hostile communities.
To do this, we collected for each emitting subreddit:
- the total number of emitted links (
n_links), - the number of positive and negative links,
- the average sentiment score (
conflict_ratio).
We then ranked the subreddits according to their activity and their conflict ratio in order to identify interaction hubs as well as their behavior.
Top 10 most active subreddits (number of emitted links):
subredditdrama: 24,932 links (35% negative)bestof: 21,128 links (84% positive)titlegore,shitpost,circlebroke2
Top 10 most hostile subreddits:
fuckredditmods,mongolianhatewatch(only negative links)
Top 10 most cooperative subreddits:
episodehub,ipad,norwayonreddit(only positive links)
Descriptive statistics on the conflict_ratio (all subreddits):
- Mean: 0.91, Median: 1.0
This initial exploration suggests several structural characteristics of the social ecosystem that will need to be analyzed in more detail in our upcoming analyses:
- A large share of interactions seems to come from a minority of larger communities (
subredditdramaalone responsible for 24,932 links, more than 48% of total emitted links). - The overall nature of interactions is very benevolent, but a small proportion of small subreddits are exclusively belligerent.
The highly asymmetric distribution of interactions suggests the presence of hubs concentrating interactions and influencing global dynamics through their behavior and activity. This asymmetry, as well as that of behaviors, also raises the question of the potential reciprocity of these dynamics. These two suggestions will be explored through our next analyses.
In order to explore the asymmetric distribution of the nature of interactions suggested by our previous analyses, we wanted to examine whether communities tend to adopt a consistent behavior in all their interactions or if this behavior changes.
For this, we studied the distribution of the conflict_ratio of all communities using a histogram.
As we could expect from our first analysis suggesting an asymmetric distribution of conflict_ratio, we had to use a logarithmic scale on the y-axis.
Thanks to this histogram, we confirm that a large number of communities have either mostly benevolent or mostly conflictual interactions. The social ecosystem of Reddit is therefore strongly polarized in the nature of its interactions and contains both belligerent and friendly hubs, whose influence will be analyzed in our upcoming sections.
We have seen that some communities seem to play a central role in the exchanges and that their behaviour can potentially fuel or reduce global conflicts. It would now be interesting to examine the nature and number of their interactions, to identify them as well as their behaviour.
We selected the 25 most active emitting and receiving communities to more precisely observe their interactions through two complementary heatmaps:
- The first represents the total number of unidirectional links between each pair of these communities
- The second indicates the
conflict_ratiobetween these same pairs.
This double visualization allows us to see, in central communities, who interacts with whom and with what behaviour.
Beyond the identification of hubs already established previously, the first heatmap shows us that emitter hubs are not the same as receiver hubs.
There is therefore a distribution of roles, with certain communities emitting and others particularly receiving.
As for the second, it shows a great heterogeneity in the nature of interactions between pairs. More precisely:
- some belligerent communities like shitstatistessay or srssucks have very negative interactions
- others, like bestof or switcharoo, have mostly peaceful interactions
- moreover, we see more aggressive behaviour between communities with opposing ideologies
One last interesting element to observe is that the behaviour of interactions seems more homogeneous in outgoing interactions than in incoming ones.
This suggests that communities adopt a consistent stance in their interactions with others, while the behaviour adopted by other communities toward the same one is very variable.
In other words, communities are consistent in their behaviour toward others but are not perceived in the same way by other communities.
These analyses confirm the existence of large communities with aggressive and friendly tendencies and raise the hypothesis of dominating dynamics and polarization.
These hypotheses will be studied in the next analyses, particularly in the ones on reciprocity and thematic correlations.
Several of our previous analyses have revealed central communities in the interactions. We will now make a global visualization of all communities in order to have an overall view of this phenomenon.
A histogram makes it possible to see this global distribution and is made optimal by a log-log visualization.
The graph confirms that the majority of communities have very few interactions, while a very small number have an enormous amount.
This characterization is in fact typical of social ecosystems and is called a scale-free structure. This comes from the fact that interactions with large communities are facilitated by their accessibility (attachment preference), whereas less frequented communities are harder to reach. In other words, the more a community is known, the more visible it is. Conversely, the more niche a community is, the less it will be highlighted.
This observation reconfirms the unequal structure of this unequal social ecosystem.
We can now ask to what extent this influences the nature and intensity of the global interactions of the ecosystem, which will be analyzed in our next points on the reciprocity of exchanges.
Our previous analyses revealed that interactions on Reddit are highly asymmetric and often centralized around a few influential communities. To refine our understanding of these hubs, we now examine a more subtle question: how does reciprocity vary with the size of a community and the intensity of its connections?
To investigate this, we conducted two complementary analyses:
- Subreddit-level: We measured the ratio of emitted links over total links (emitted + received) for each community, then grouped them by their out-degree (total number of emitted links).
- Pair-level: We computed this same ratio for each subreddit pair (A → B vs B → A), grouped by the total number of links exchanged between them.
Given the vast disparity in community sizes and interaction volumes, we applied logarithmic binning to better capture the trends. For each bin, we calculated the mean reciprocity ratio and standard deviation.
The first graph reveals a slight decrease in average reciprocity as subreddit activity increases. Larger subreddits tend to exhibit a more unilateral behavior, but this effect remains relatively weak. This suggests that size alone is not a major determinant of reciprocity.
In contrast, the second graph shows a much sharper drop in reciprocity as the intensity of interaction between subreddit pairs increases. When two communities interact frequently, their exchanges tend to become more asymmetric, with one subreddit increasingly dominating the flow of mentions. Interestingly, this trend slightly reverses in the densest pairs, where mutual engagement appears to resurface.
These results suggest that unilateral dynamics are not necessarily driven by size, but rather by the depth of bilateral relationships. The more two communities interact, the more likely one is to take the lead—whether for support, provocation, or dominance.
In short, Reddit’s social engine is not just shaped by a few large players, but by how deep and directed their interactions become. Reciprocity fades not with fame, but with focus.
We will now turn to a previously mentioned hypothesis reflecting a common behavior that fuels belligerent interactions between communities. It is the question of reciprocity in conflictual interactions at the level of the community and the relationship between two communities. More precisely, on the one hand, is the level of belligerent interaction suffered by a community correlated with its behavior towards other communities? And on the other hand, does a belligerent interaction from one community towards another trigger a response in return?
To observe this, it is enough to compare the conflict_ratio emitted and received by a community on two scales:
- A first scale at the level of subreddits used to investigate a possible symmetry between the attitude of a community towards others and the way it is perceived in return
- And a second scale at the level of pairs of communities allowing us to see if the attack of a community A towards a community B will trigger a response from B to A
We also placed on the graph the line y = x illustrating a perfectly reciprocal hostility between two communities for a visually expressive rendering.
The first graph shows that overall, the more a community is aggressive with others, the more other communities are aggressive towards it in return. However, this is not true for communities with extreme behaviors. Indeed, even if a community is very pacifist, it will still suffer attacks, and if a community is extremely belligerent, it will be proportionally less punished in return.
The second graph confirms the presence of a slight global correlation that also applies at the level of two communities but with a more linear increase. This shows that reciprocity is more targetable at the scale of two communities than at the global scale. This makes total sense, as a relationship is more easily targetable when applied to interactions between two communities rather than one community with all the other communities. However, even if it increases, reciprocity is far from a balanced behavioral reciprocity, which is again even more true for extreme behaviors.
We thus conclude that the reciprocity of hostilities is neither systematic nor balanced.
This suggests potential dynamics of power with dominant communities (aggressive and proportionally little punished in return) and submissive communities (pacifist yet still suffering attacks),
making Reddit not only a space of exchange but also a terrain of social domination.
Reddit is not only a collection of isolated communities, but a dynamic ecosystem in which subreddits continuously interact. These interactions may express support, disagreement, or outright conflict. In this task, we move from a static view of the network to a temporal perspective, asking how cooperation and conflict evolve over time.
Our analysis focuses on the period from 2014 to 2017, during which we examine both long-term trends and short-lived fluctuations in inter-community behavior.
We begin by aggregating interactions on a yearly basis to capture structural trends while smoothing short-term variability. Two indicators are considered: the proportion of negative interactions and the average sentiment of links exchanged between subreddits.
[FIGURE 1 — Yearly evolution of negative share and mean sentiment]
(Insert yearly line plot here)
Despite year-to-year variations in activity, both indicators remain strikingly stable. Negative interactions consistently represent only a small fraction of all links, while the mean sentiment remains strongly positive throughout the period. This suggests that conflict does not gradually intensify over time, but rather coexists with a persistent cooperative backbone.
To further explore this balance, we examine how cooperative and conflictual interactions contribute to overall activity over time. Instead of focusing on absolute volumes, we track their relative proportions.
[FIGURE 2 — Stacked area chart of cooperation vs conflict]
(Insert stacked area chart here)
The dominance of cooperation is evident at all time points. Even as the total number of interactions fluctuates, the relative weight of negative links remains largely unchanged. This highlights a key structural property of Reddit’s inter-community dynamics: cooperation is not episodic, but foundational.
Yearly averages may conceal short-lived but intense episodes of conflict. To reveal these, we turn to a finer temporal resolution and analyze monthly fluctuations in negative interactions, identifying mobilization peaks using standardized z-scores.
[FIGURE 3 — Monthly mobilization peaks of negative interactions]
(Insert z-score stem / lollipop plot here)
Several sharp peaks emerge, indicating brief surges of hostility. These episodes contrast with the long-term stability observed earlier and suggest that conflict on Reddit is primarily episodic, likely triggered by specific events, rather than the result of slowly escalating tensions.
Finally, we ask whether conflict between communities tends to persist or fade quickly. For each pair of subreddits, we measure the temporal span between the first and last observed negative interaction.
[FIGURE 4 — Distribution of conflict persistence durations]
(Insert log-scale histogram here)
The distribution is heavily skewed toward short durations: most conflicts occur within narrow time windows, while long-lasting hostile relationships are rare. This further supports the view of conflict as a transient disturbance rather than a stable feature of inter-community relations.
Viewed through time, Reddit’s inter-community network appears remarkably resilient. Cooperation dominates at all scales, while conflict manifests as brief, localized bursts that fail to destabilize the system as a whole. These findings complement the structural analyses of earlier tasks and suggest that temporal dynamics play a key role in containing and dissipating conflict within online communities.