Does Reddit display brain-like behaviours? Can we “Frankenstein” Reddit?
Reddit becomes a testable brain metaphor: subreddit categories as regions, hyperlinks as wiring, and synchronized language as functional connectivity.
Executive summary
This project uses connectomics as a modelling lens, not a biological equivalence claim. The vocabulary is useful: structure, function, hubs, lesions, excitation, inhibition, and reorganization.
The output is a deployable data story: network science, NLP, sentiment, temporal correlation, and interactive visual evidence in one place.
Why it matters
38,883
category-matched hyperlinks
directed subreddit-to-subreddit links
1,900
unique subreddits
after category filtering
19
semantic regions
high-level subreddit categories
86
language features
LIWC, sentiment, readability, structure
69
continuous signals
subreddits retained for time-series analysis
2014–2017
time horizon
monthly functional-connectivity window
Choose your reading path
Interactive data anatomy
Category-matched links: The filtered directed edge set used for category-level structural and sentiment analyses.
Evidence card
Some communities co-activate over time even when they do not directly hyperlink, exposing hidden functional synchrony.
Evidence
Functional-only links remain after subtracting explicit hyperlinks from the functional connectome.
Method pipeline
Step 1
Raw Reddit hyperlinks
Step 2
Category atlas
Step 3
Structural graph
Step 4
Sentiment layers
Step 5
Monthly LIWC signals
Step 6
Functional connectome
Step 7
Brain-space projection
Zoom out from Reddit and individual posts blur into communities exchanging signals. Some links are explicit wiring. Other relationships only appear when communities rise and fall together over time.
The analogy is compact: subreddits are neurons, hyperlinks are structural connections, and correlated LIWC signals are functional connections. The site asks whether that abstraction reveals patterns that are hard to see in the raw hyperlink graph.
To make the map interpretable, we use semantic categories as an atlas. They do not cover every subreddit, but they provide stable landmarks for studying structure, sentiment, lesions, and rhythm.
Here, a category means a high-level label (e.g., “Technology”, “Sports & Games”, “News & Politics”) assigned to a subreddit for analysis.
Our analogy (simple, but useful)
Crash course: neuroscience tools we “steal” (click to jump)
Atlas building
Do our labeled neurons represent the whole brain?
Neuron types
Sensory / interneuron / motor roles from wiring patterns.
Positive & negative circuits
Excitatory vs inhibitory flows across categories.
Brain trauma
Neuroplasticity after a subreddit ban (lesion).
MRI / fMRI
Structural connectome vs functional co-activation networks.
Firing dynamics
Temporal co-activation and animated connectivity.
Question 1
Do subreddit categories exhibit distinct functional roles (sensory / interneuron / motor) based on in-degree and out-degree?
Question 2
Can we define a functional network of subreddits beyond explicit structural hyperlinks (via temporal co-activation)?
Question 3
To what extent does functional connectivity transcend structural wiring, and what characterizes these remotely communicating communities?
Interactives: controls
Zoom. Open plots in a larger focused view.
Slider. Drag the white knob to move through time.
Snapshot. Each slider position updates the visual state.
Filters. Toggle positive, negative, or combined edges.
The category atlas only labels part of Reddit. Before using it as a map, we test whether those labeled communities preserve the dominant behavior of the full network.
The short answer: the subset captures the core distributional structure, while sacrificing rare long-tail communities.
Each hyperlink carries a linguistic fingerprint: length, structure, sentiment, and LIWC-derived psychological markers. These features describe how communities communicate, not only who connects to whom.
We compare the labeled subset with the full network through feature distributions. Their central tendencies and shapes largely overlap.
That makes the labeled communities useful landmarks: not exhaustive, but typical enough to orient the map.
The lost portion is mostly the edge of the feature space: rare styles, niche communities, and low-frequency behavior. The tradeoff is explicit: we gain interpretability while losing some long-tail specificity.
PCA is fitted on the full dataset, then both the full network and labeled subset are projected into the same state space. The subset occupies the same core regions along the leading components.
In practical terms, it preserves the main modes of variation while smoothing away idiosyncratic extremes.
The labeled subset works as a valid working atlas: not complete, but representative enough to assign regions and study structure–function relationships.
We project subreddit categories onto an MNI brain template through thematic clusters: groups of categories mapped to semantic concepts inspired by the Huth et al. framework.
The result is not a literal brain claim; it is an interpretable spatial scaffold for comparing communities, links, and functional co-activation.
The structural network is built from explicit hyperlinks: who links to whom. We ask whether categories form recognizable wiring roles: receivers, broadcasters, relays, and long-range connectors.
Figure 3 – Category in vs out degree
Figure 4 – “Axon lengths” via embedding distance
Neuron types (framing)
The analogy is role-based: incoming links resemble sensory input, outgoing links resemble motor output, and balanced categories behave like relays.
Positive and neutral links act as supportive signals; negative links act as antagonistic signals. We use this as a simple excitation / inhibition layer over the hyperlink graph.
The question becomes: who broadcasts support, who broadcasts conflict, and which communities absorb the most negativity?
Figure 8 – Sentiment received by categories
Figure 9 – Sentiment sent by categories
Interactive – Positive/Negative visualization
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
In this visualization, we examine the axons projecting from every source category. Source categories can be selected from the menu.
Treat this as a conflict atlas: red arrows and warm tones expose where sentiment becomes antagonistic.
Embedding distance becomes a proxy for axon length: short distances connect similar audiences, while long distances bridge distant communities.
The interactive “explodes” the graph into parallel connections so distance, sentiment, and direction can be inspected category by category.
Interactive – Distance comparisons
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
Use the explorer to spot long-distance links: categories that connect beyond their immediate semantic neighborhood.
Direction also matters. Some categories send many links to a target but receive very few in return, revealing asymmetric attention.
The key test is whether longer semantic jumps are associated with more negative sentiment.
For Technology, positivity drops strongly as embedding distance increases, suggesting that distant audiences are more likely to receive antagonistic links.
Finally, we test hub dependence: does the network survive random damage, or collapse when the biggest hubs are removed first?
The grey curve removes nodes randomly. The colored curve removes highest-degree nodes first.
A steeper colored drop means the layer depends on a few hubs.
Interactive – Positive hub visualization
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
Interactive – Negative hub visualization
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
In the positive network, the largest hubs are Business, Finance and Crypto, Funny/Memes, and Technology. Being a positive hub has nothing to do with sentiment proportion; rather, through sheer volume of interactions, these categories keep the positive network connected.
In the negative network, Funny/Memes, Technology, and News and Politics dominate. The red curve drops sharply: negativity is more hub-dependent than positivity.
Takeaway: large-scale negativity is not evenly distributed; it is amplified by a small number of influential hubs.
On June 10, 2015, Reddit banned five large subreddits. We treat this as a network lesion: a sudden removal of major hubsfollowed by measurable shifts in platform-wide language.
The finding is not just local removal. After the ban, aggregate discourse becomes more negative, more certain, more out-group focused, and more linguistically complex.
Daily aggregation controls for changing post volume and lets us compare affect, cognition, social orientation, and discourse style before vs after the event.
MANOVA tests whether the joint psychological profile changes after the ban, rather than testing each feature in isolation.
The MANOVA revealed a significant systemic change in the network’s psychological profile following the ban (Pillai’s trace = 0.18, F(6, 1210) = 45.31, p < .001).
Welch tests then identify which features move and in what direction.
| Feature | Mean Change (After − Before) | p-value | Direction | Interpretation | Collective Meaning |
|---|---|---|---|---|---|
| VADER_Compound | −0.0159 | < 1e-17 | ↓ | Overall sentiment became more negative | Increased emotional negativity |
| LIWC_Negemo | +0.00063 | < 1e-10 | ↑ | More negative emotional language | Heightened emotional intensity |
| LIWC_Anger | +0.00039 | < 1e-7 | ↑ | Increased expressions of anger | Grievance and hostility |
| LIWC_Anx | +0.00008 | 0.0013 | ↑ | More anxiety-related language | Perceived threat / uncertainty |
| LIWC_CogMech | −0.00239 | < 1e-20 | ↓ | Reduced cognitive processing language | Less deliberative reasoning |
| LIWC_Insight | −0.00116 | < 1e-35 | ↓ | Fewer reflective / insight terms | Reduced self-reflection |
| LIWC_Tentat | −0.00093 | < 1e-28 | ↓ | Less tentative language | Decreased nuance / openness |
| LIWC_Certain | +0.00021 | < 1e-4 | ↑ | More certainty and absolutist framing | Cognitive rigidity |
| LIWC_They | +0.00026 | < 1e-13 | ↑ | Increased out-group references | Stronger us-vs-them framing |
| LIWC_You | −0.00019 | < 1e-4 | ↓ | Fewer direct addresses | Reduced interpersonal engagement |
| LIWC_We | −0.00003 | 0.58 (ns) | ≈ | No change in inclusive language | No increase in group solidarity |
| Readability_index | +0.48 | < 1e-5 | ↑ | Higher linguistic complexity | More argumentative discourse |
| Avg_words_per_sentence | +0.84 | < 1e-4 | ↑ | Longer sentences | Increased rhetorical elaboration |
Sentiment and affect. Overall sentiment became more negative, with decreases in VADER Compound scores and increases in negative emotion and anger, indicating a more emotionally charged discourse.
Cognitive style. Insight, cognitive processing, and tentative language decreased, while certainty increased, reflecting reduced deliberation and greater cognitive rigidity.
Social orientation. References to out-groups (They) increased, direct address (You) decreased, and inclusive language (We) showed no significant change, suggesting a shift toward adversarial framing without increased group cohesion.
Discourse complexity. Readability and average sentence length increased, indicating more elaborated and argumentative posts.
Collectively, these patterns suggest that the ban functioned as a coordinated disruption in the network’s social brain.
Conclusion
Removing major hubs was associated with systemic changes in emotional, cognitive, and social language patterns.
Causality remains bounded: the timing is consistent with a network-level disruption, but concurrent external factors cannot be fully excluded.
Structural connectivity asks who links to whom. Functional connectivity asks which communities move together over time. The key question is whether Reddit contains synchrony beyond explicit hyperlinks.
Interactive – Functional connectivity difference
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
Turning links into neural signals
A hyperlink is only the wiring. The signal lives in the post behind it: sentiment, LIWC features, timing, and linguistic style.
We retain subreddits that fire continuously enough to support time-series analysis, then aggregate activity into monthly bins.
The selection balances two constraints:
We then choose the smallest time bin that preserves the most continuous signals.
CONNECTIVITY FILTER – MOST SYNAPTICALLY ACTIVE SUBREDDITS
Bin selection – continuity vs coverage
We rank subreddits by total hyperlink volume (incoming + outgoing) and keep the most synaptically active communities, ensuring our time-series analysis focuses on neurons that keep firing.
Result
The winner is 1 month: stable, continuous, and still granular.
The final issue is dimensionality: 86 features per subreddit are too fragmented for a readable functional map.
We compress related LIWC variables into six functional signals. These become the activation traces used to build the functional connectome.
Interactive – Cluster time-series explorer
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
Interactive – Functional connectivity
Interactive view ready on demand
Click to load the full interactive; brain maps include legend and hover focus.
The punchline: wiring vs rhythm
We compute temporal correlations across the six functional signals and ask: who moves together even without direct links?
From subreddit topology to brain space
The comparison separates two views: all significant functional links, and functional minus structural connectivity. The second view isolates communities that synchronize without directly linking.
This answers the central question: is Reddit only explicit interaction, or does it also contain hidden shared rhythm?
The punchline: subtracting hyperlinks reveals who moves together anyway.
We combine multiple public datasets and several preprocessing steps to construct the Reddit brain. Below is a high-level overview of the pipeline that underlies all the visualizations above.
Important framing
This is a computational analogy, not a biological claim. Brain terms are used to structure the analysis: they help compare wiring, co-activation, hubs, and perturbations in a large social network.
All figures on this page are generated from our analysis notebooks and exported as static assets for this front-end.
What I built
Limitations
Continue exploring
The repository contains the notebooks, preprocessing scripts, static assets, regression tests, and GitHub Pages workflow behind this site.