Conflict-monitoring datasets (like ACLED or GDELT) log thousands of political-violence events per day. Most days look statistically ordinary; a small number are extreme outliers — sudden surges in event count or fatalities that deserve a human analyst's attention. Isolation Forest finds those outliers without any labels: it builds many random binary-split trees over the data's features, and points that get separated ("isolated") from the rest in very few splits are scored as anomalous.
IForest exposes via n_estimators.Isolation Forest (Liu, Ting & Zhou, 2008) is unusual among anomaly detectors because it never builds a profile of "normal" — it directly exploits the fact that anomalies are "few and different," which makes it fast and cheap to run over large open conflict-event archives before any human review begins.
A synthetic daily conflict-event log is rendered as a 3D bar chart, and a real, simplified Isolation Forest runs in the browser to score every day for anomaly, flagging the statistically unusual spikes for review.
Isolation Forest repeatedly partitions the data with random splits; points that separate from the rest in very few splits get short average path lengths and high anomaly scores — exactly what is animated by the translucent split planes.
Adjust contamination to control how many top-scoring days get flagged, change tree count to see the score stabilize, and inject spikes to watch the model react. Toggle the split-plane animation and regenerate the series any time.
Because Isolation Forest needs no labelled anomalies to train on, it's a common first pass in open-data pipelines like ACLED or GDELT before analysts manually review flagged events.