mdpi-hypernetwork-archetype

Compute per-user mean engagement score, VADER compound sentiment, and Detoxify toxicity, normalize each axis to [0, 1], then split each axis at a threshold (0.5 or corpus median) into High/Low. The three labels combine into an archetype like HHL or LLH.

Sentiment measures affective valence (positive vs negative tone) while toxicity measures probability of harmful language. They are independent: a comment can be positive but toxic ("Hell yeah, that's awesome!") or negative but civil ("I respectfully disagree").

Detoxify (unitary/toxic-bert) is the preferred option, providing multilabel toxicity probabilities in [0, 1] without rate limits. Perspective API also outputs [0, 1] probabilities but is rate-limited. Both require documenting the model version for reproducibility.

At least 30 users are required; below that, report individual axis scores without archetype assignment. Corpora of 100-499 users are adequate, while 500+ users support full analysis with all eight archetypes meaningfully populated.

Fall back to two-axis classification on a 2x2 grid, such as Score x Sentiment when toxicity is unavailable. Document which axis was missing and which archetype distinctions could not be made.

Archetypes describe behavioral patterns observed in a specific time period, not stable traits or identities. Users can shift archetypes over time, and small threshold changes can reassign boundary users, so labels are snapshots rather than identities.