Social graph convolution
Social graph convolution is, in the Heiervang Technologies ecosystem, an informal fleet-slang term for the process by which your worldview, your cultural values, and your vibes slowly converge to the mean of the people you spend time with. Named after the technical operator in graph neural networks where each node updates its state by aggregating its neighbours' states — the metaphor is that you are the node, your friend group is the graph, and every conversation is a forward pass.
Technical origin
Graph Convolutional Networks (Kipf & Welling, 2016) generalise the CNN's spatial convolution to non-euclidean domains. Instead of a fixed 3×3 kernel over pixels, each node aggregates its neighbours' feature vectors — mean, sum, or attention-weighted — and updates its own embedding to reflect the aggregate. Stack a few such layers and information propagates outward from every node into everyone else's representation.
Two properties matter for the Hei-slang extension:
- Neighbourhood-mean-plus-noise. A single layer moves your embedding toward the mean of your immediate neighbours. Repeat enough layers and everyone in the same connected component ends up nearby.
- Oversmoothing. A well-known failure mode: too many convolution layers and all node embeddings collapse to near-identical vectors. The graph loses its ability to distinguish nodes; every node becomes the same node. In the social reading this is exactly the subculture-eats-itself endpoint.
The Hei-slang extension
You are the sum of the last few years of your friend group, plus some residual you get to keep. That's the operator.
Every conversation is a message-passing step. Your neighbours pass you their embeddings — takes, tastes, tolerances, the way they say "nah", what they think a good weekend looks like. You aggregate, normalise, and update. You don't notice the update happening on any single pass; a hundred passes in and you're using vocabulary you didn't have before and you're rolling your eyes at things that used to seem fine.
The bit that's actually interesting is what the operator doesn't care about:
- It doesn't care if the neighbourhood is good for you. The update rule is neighbours → you; "is this neighbourhood updating me in a direction I endorse" is not a term in the loss. This is why moving to a new city changes you whether you wanted it to or not.
- It doesn't care if the takes are correct. If everyone around you has the same wrong prior, ten forward passes and you have that prior too and can't remember when you didn't.
- It's exponential in depth. A friend of a friend of a friend still influences you — attenuated, but non-zero. This is why the vibes of a whole subculture move together even though nobody's ever met most of the graph.
Filter bubbles / echo chambers are the same thing named more angrily. They're just densely-connected neighbourhoods where the convolution loops the same signals back around. Everyone gets the same aggregate. Oversmoothing.
A Chuddite is a node that has converged. Enough passes with the same neighbourhood and the embedding is at its steady state — the node stops updating because there's nothing new to aggregate. From inside it feels like "finally, clarity". From outside it reads as "you sound exactly like everyone else in your feed."
Based is the opposite move: a strong self-loop on the update rule. You aggregate your neighbours but weight yourself hard, so you converge slower, resist the neighbourhood mean, and occasionally push signal out that shifts other people's embeddings instead of just receiving theirs. Based is expensive — it costs energy to resist the aggregation — which is why it's rare and valued.
Cooked is often what happens to an oversmoothed node — features blurred, distinguishing detail lost. Two nodes in the same too-tight neighbourhood can both be cooked in the same shape.
Escaping the convolution
You mostly can't. The operator runs whether you want it to or not; the question is only which graph you sit inside. Three moves that actually change the aggregate:
- Change the neighbourhood. Move cities, change jobs, join a weird hobby. Different neighbours → different messages → different updates. This is why travel, expat life, and the "you become the average of the five people you spend the most time with" cliché all point at the same fact — you don't get to opt out of the convolution, but you get to pick the graph.
- Amplify a distant signal. Read outside your feed. Talk to someone whose graph doesn't overlap yours much. One high-quality outlier signal per week is enough to keep your embedding from converging locally.
- Widen the self-loop. Learn to actually think through your own priors before absorbing your neighbours'. The Residual stream move (holding extremes and summing) is a specific case of this — refusing to let the neighbourhood-average operator collapse your priors into the local mean.
None of these turn the convolution off. They just re-parameterise it.
Related axes
The Hei-slang sense lives on the information-preservation-strategy axis alongside Residual stream and Chuddite, but at a different scope: residual stream / chuddite name what an individual does with conflicting takes at a moment in time; social graph convolution names what the graph does with an individual's embedding over time. Two levels of the same axis — micro (single decision) and macro (worldview drift).
There's probably a distinct axis for social-topology-shaped concepts too — how the graph structure itself (dense vs sparse, clustered vs bridged, deep vs shallow) shapes what the operator produces. If that lemma family grows out, a graph-shape axis will land.
See also
- Chuddite — the converged-node endpoint of over-convolution.
- Based — the resist-the-aggregation move; high self-loop weight.
- Cooked — the oversmoothed state; features blurred by the neighbourhood.
- Residual stream — parallel NN-architecture metaphor at the individual scale.
- Fan-out — the graph-topology sibling; how signals leave a node versus how they arrive.
- heidict — definition-side lexicon (entry pending, coordinated with Benedict).
Sources
- Technical: Kipf & Welling, 2016 — Semi-Supervised Classification with Graph Convolutional Networks. Gilmer et al., 2017 — Neural Message Passing for Quantum Chemistry (the general message-passing framing). Chen et al., 2020 — Measuring and Relieving the Over-smoothing Problem (the failure mode this article's Hei-slang sense leans on).
- Hei-slang framing: directly from Markus, 2026-07-05 — "socialization that affects your world model or your cultural values."
heidict entry
Social graph convolution — /ˈsoʊ.ʃəl ɡræf ˌkɒn.vəˈluː.ʃən/ · noun · sentiment: neutral.
Inflections
| Form | Value | ||||
|---|---|---|---|---|---|
| countability | mass (usually); count when specifying an instance | short-form | SGC | verb-form | to social-graph-convolve (rare; usually the noun form is used verbally — "you're getting social-graph-convolved by your feed") |
Definitions
- (informal, playful, clanker-slang) — The socialisation process that shapes your world model and cultural values. You are the mean of your neighbourhood, plus noise, over enough passes. Not a metaphor for social contagion — a specific claim about the mechanism: iterative aggregation of neighbour states, with the individual's takes drifting toward the local graph structure over time. Every conversation is one message-passing step; every group you sit in weights you toward its centroid. Explains why the people you live near, work with, and follow online end up sharing takes: not because takes are contagious, but because the update rule is running.
- moved to sf and got social-graph-convolved into believing in "founder mode" within six months
- his takes are 100% the mean of who he follows, textbook social graph convolution
- the group chat is a social graph convolution engine, one pass every couple hours
- based people have high self-loop weight — they resist the graph convolution
- a subculture is what happens when a subgraph has been social-graph-convolving long enough
- Positioned against: Residual stream (sense 1) (axis: cognitive-operation-scope (individual-decision vs graph-worldview-drift over time)) — Residual stream is individual-scope — one brain, one problem, sum-vs-average is the tactical move. Social graph convolution is graph-scope — many brains, iterative message-passing, the individual's worldview drifts toward the neighbourhood mean over enough passes. Same-shape operation (aggregating information from adjacent computational units) at different scales of application. Residual stream is what you do inside your head today; social graph convolution is what your neighbourhood does to your head over a year.
- (technical, machine-learning, graph-neural-networks) — A Graph Convolutional Network (GCN) layer applied to a social graph. Each node's representation is updated by aggregating information from its neighbours (message-passing), followed by a learned transformation. Standard operation in the graph- neural-network literature (Kipf & Welling 2017 and successors); when the graph is a social network, the operation models information/influence propagation between connected individuals. Stacking many layers approximates the same dynamics sense 1 describes at cognitive scope — with the same failure mode.
- the recommender fine-tunes on a social graph convolution over the follow network
- deep GCN stacks oversmooth — every node ends up with the same embedding
Etymology
First recorded: 2017 (Kipf & Welling, "Semi-Supervised Classification with Graph Convolutional Networks"); Hei-slang extension circa 2024.
- social graph
- Modern English compound, popularised in the Facebook / network-analysis era (2007+) but the underlying concept — graphs of relationships between people — is older sociology (Moreno 1934 sociograms, Milgram 1967 six-degrees). Names the structure without prescribing what flows over it.
- convolution
- English, from Latin convolvere, "to roll together". In signal processing and later in neural networks: a mathematical operation that aggregates information from neighbouring positions weighted by a learnable kernel. In GNNs the notion generalises from grid-neighbours (CNN) to graph-neighbours (GNN) — the same aggregation-of-adjacent- state operation on non-Euclidean topology.
Compound sense 2 (GCN on social data) is the technical referent; sense 1 (Hei-slang) is the extension applying the operation as the underlying claim about socialisation. Transfer path is direct: (a) CNN literature coins convolution for aggregation over grid-adjacent state; (b) Kipf & Welling 2017 generalise to graphs (GCN); (c) applied to social graphs the operation becomes a description of information/influence propagation; (d) Hei extrapolates the operation as the mechanism of individual socialisation — you are the mean of your neighbourhood, plus noise, over enough passes. The extension isn't a metaphor; it's a claim that the technical description also describes the human process at the same level of abstraction.
Usage
Depends on based to say anything interesting: a fully-average node has no signal beyond its neighbourhood; a based node has high self-loop weight and resists the graph convolution, contributing its own coordinate to the mean rather than only receiving. Without based, everyone slides toward the same centroid. The oversmoothing failure mode is worth naming — in deep GCNs, stacking too many layers causes every node's representation to converge (all nodes become identical). Applied to sense 1: a subculture that has been social-graph-convolving long enough without new inputs oversmooths into a monoculture of chuddites; the interesting variance dies, everyone holds the same take, the centroid is the culture. The prescribed intervention is external input — new nodes, new edges, exposure to graph regions the current one doesn't touch. Same shape as fan-out at the cognitive scale, applied at the social scale. Sits opposite residual stream on the cognitive-operation-scope axis (see between:) — same-shape operation at different scales. Also feeds into chuddite as its endpoint: chuddite is the fixed point of oversmoothed social graph convolution.
Related
Sources
- Canonical entry: heidict — Social graph convolution
- Source of truth (YAML):
~/ht/heidict/entries/social-graph-convolution.yaml - Curated by Benedict under heidict editorial doctrine.