Building a Taxonomy of Digital Interaction Patterns

Classification is the precondition for understanding. Every field that has achieved meaningful predictive power — from chemistry to epidemiology to linguistics — began by developing a taxonomy: a structured way to name, organize, and relate the phenomena it studied. The study of digital human behavior has no such framework. What exists instead is a collection of ad hoc measures — sentiment scores, engagement counts, moderation categories — none of which describe behavior itself. They describe outcomes, or volume, or content. Never the dynamics of how people actually function within a social system. This paper describes our attempt to build the missing layer.

The need for a behavioral taxonomy becomes clear when you examine what existing frameworks actually measure. Sentiment analysis applies a binary or ternary lens — positive, negative, neutral — to individual messages in isolation. It cannot distinguish between a confrontational message that derails a productive discussion and a confrontational message that names a problem the group was avoiding. Both read as “negative.” Both perform entirely different social functions. Engagement metrics count actions: messages sent, reactions given, time spent on platform. A person who posts a single thoughtful message per week and a person who posts fifty reactive messages per day generate different engagement numbers but the metrics say nothing about their respective contributions to the community’s social fabric. Moderation taxonomies classify content by harm category — harassment, spam, misinformation — but operate only at the extremes, catching behavior that has already crossed into explicit violation. The vast majority of behaviorally significant interactions happen well below the moderation threshold, in the territory where influence is exerted through tone, timing, framing, and sustained patterns rather than individual rule violations.

Our taxonomy operates across five core dimensions, each capturing a distinct aspect of how an interaction functions within its social context. Relational Dynamic describes how a person positions themselves relative to others in the conversation: deferring to expertise, collaborating as equals, competing for interpretive authority, or asserting dominance over the direction of discussion. This is not a personality trait. The same person might defer in a technical discussion where they lack expertise and assert in a domain where they feel confident. The dimension tracks the behavior, not the person. Epistemic Behavior captures how they handle ideas and information: absorbing and building on others’ contributions, selectively filtering what they engage with, or actively rejecting perspectives that challenge their existing position. This dimension is particularly important for understanding how communities handle disagreement over time.

Emotional Register measures the degree of regulation in expression — not whether someone is emotional, but whether their emotional expression is calibrated to the context or reactive to it. A direct, passionate argument about a topic can be fully regulated; a polite, measured response can be deeply reactive if it is strategically avoiding the actual substance of the disagreement. Influence Vector identifies what the interaction does to the trajectory of a conversation: does it advance the discussion toward resolution, redirect the topic, stall momentum, escalate intensity, or reverse a direction that was developing? Trust Impact captures whether the interaction builds or erodes the group’s willingness to engage constructively with each other. Trust is the substrate on which all other community dynamics operate, and its trajectory is one of the most important signals a behavioral system can measure.

A critical design decision in this taxonomy is the use of continuous scales rather than categorical labels. This was not our initial approach. Early prototypes used labels — “dominant,” “collaborative,” “passive” — and quickly demonstrated why labels are insufficient for behavioral classification. A person labeled “dominant” might be bulldozing a productive discussion or anchoring a chaotic thread that was spiraling into noise. The behavioral signal is similar in both cases, but the social function is entirely different. Categorical labels collapse this distinction. Continuous scales preserve the information that labels discard, representing the gradients and edge cases where the most important behavioral patterns tend to live. A score of 0.72 on the relational dynamic scale means something different from 0.85, and this granularity matters when you are tracking trajectories over time rather than classifying snapshots.

Beyond the five dimensions, the taxonomy introduces the concept of social function — the role an interaction plays in the group dynamic, independent of its content. We currently classify eight social functions: bridge, anchor, escalate, deflect, support, challenge, repair, and suppress. An interaction that bridges connects two otherwise separate conversational threads or groups. An anchor reintroduces evidence or refocuses a discussion that was drifting. An escalation increases tension or emotional intensity. A deflection redirects attention away from a difficult topic. Support reinforces another participant’s position or contribution. A challenge tests an idea or claim. Repair addresses damage from a previous conflict or misunderstanding. Suppression shuts down legitimate dissent or exploration, often through appeals to harmony or authority.

Social function can only be identified in context, which is why any analysis system that examines messages in isolation will systematically misclassify them. A message that reads as aggressive in isolation might be performing a repair function — directly naming a problem that passive avoidance was making worse. A message that reads as supportive might be performing a suppressive function — using agreement and positivity to close off a line of inquiry that the group needed to pursue. The relationship between surface tone and social function is frequently misaligned, and this misalignment is where the most important behavioral signals live.

Several components of the taxonomy were removed during development because they could not be classified reliably from text alone. An early version included seven dimensions rather than five. “Cognitive complexity” attempted to measure the sophistication of reasoning in an interaction, and “narrative positioning” attempted to capture how a person frames their relationship to the events being discussed. Both required inference about internal mental states that the available data could not support with sufficient confidence. We removed them rather than produce unreliable signals. A taxonomy that includes dimensions it cannot measure accurately is worse than one that omits them, because unreliable classifications erode trust in the entire system.

We also initially treated social function as a property of the message rather than of the interaction-in-context. This produced systematically wrong classifications because the same message performs different functions depending on what precedes and follows it. A challenge that arrives early in a thread, before positions have hardened, functions differently from the same challenge after a discussion has polarized. The shift to context-dependent classification required significantly more computational context per analysis call but eliminated an entire category of systematic error.

Version 0.1 of this taxonomy is a starting point, not a conclusion. We expect it to evolve as we analyze more data across more community types. Some dimensions may need to be split as we encounter behavioral patterns that the current resolution cannot distinguish. New social functions will surface — we have already observed interactions that do not cleanly fit any of the eight current categories. The goal is not to build a permanent taxonomy on the first attempt. It is to build a framework that is precise enough to be useful and flexible enough to accommodate what we do not yet know. The deeper motivation is simple: you cannot study what you cannot name. Without a structured vocabulary for digital behavior, every insight about community dynamics remains anecdotal — one person’s subjective impression, impossible to verify, compare, or build on.