1. Traits
Each node starts with a small conversational profile. Those features help, but they are not enough on their own.
Explorable essay · synthetic interaction graphs
A person can look fine in isolation and still get overloaded in a group. The difference is in the links: who interrupts, who redirects, who gets reinforced, and who keeps absorbing the spillover. This experiment trains a small graph neural network on that structure.
Each participant has a small set of conversational tendencies, but the label is decided by the surrounding graph: incoming pressure, repair attempts, and whether the most dominant speaker is being amplified by other people in the room.
The examples below are intentionally cascade-heavy, so the group structure is easy to see at a glance. A trait-only baseline can read a participant in isolation. The GNN can also read the neighborhood around them.
Nodes encode conversational preferences such as expressiveness, reciprocity, processing need, and resilience. Directed edges are built from signals such as airtime pressure, interruption rate, repair support, redirect pressure, and long-turn intensity.
Here, pressure means someone repeatedly pushes attention onto another person. Repair means attempts to soften, steady, or redirect a rough exchange.
The label is a synthetic estimate of who tips into overload once those interactions accumulate. This explainer focuses on rooms where that reinforcement spreads broadly enough to be visible, rather than on narrow edge cases.
Each node starts with a small conversational profile. Those features help, but they are not enough on their own.
Pressure and repair travel along directed links. The same person can be stable in one neighborhood and overloaded in another.
The GNN helps when overload depends on who is being backed by the rest of the conversation graph.
Figure 1
Loading case…
01
The first view marks the final threshold directly. These rooms are cascade-heavy, so several participants may already be over the line; the useful comparison is who the trait-only baseline still misses.
02
Pressure is not evenly distributed. One participant tends to hold the floor and push attention outward, while another keeps absorbing the spillover.
03
This view only sees the node profile. It can spot vulnerability, but it misses how the room amplifies pressure through repeated interaction.
04
This final view is best read as a comparison, not as an explanation. The clearest difference usually appears where the two models disagree about the same participant.
In easy cases, both models agree. The useful cases are the ones where the trait-only baseline hesitates and the graph model does not.
Those are usually cases where pressure is being reinforced by the room, not just delivered by one speaker. The graph model can follow that pattern. The trait-only baseline cannot.
Figure 2
X-axis: training epoch. Y-axis: validation F1.
Case archive
These held-out rooms are intentionally cascade-heavy, so the structure and model differences are easy to inspect.
The experiment stays small on purpose. You can see the nodes, trace the strongest pressure and repair links, click the participants, and inspect how much pressure lands on each person. That makes the model easier to reason about.
The natural next step is a graph transformer on the same task: keep the same synthetic world and compare message passing with global attention.