Isolation Forest may look technical, but its idea is simple: isolate points using random splits. If a point is isolated quickly, it is an anomaly; if it takes many splits, it is normal.
Using the tiny dataset 1, 2, 3, 9, we can see the logic clearly. We build several random trees, measure how many splits each point needs, average the depths, and convert them into anomaly scores. Short depths become scores close to 1, long depths close to 0.
The Excel implementation is painful, but the algorithm itself is elegant. It scales to many features, makes no assumptions about distributions, and even works with categorical data. Above all, Isolation Forest asks a different question: not “What is normal?”, but “How fast can I isolate this point?”
The post The Machine Learning “Advent Calendar” Day 8: Isolation Forest in Excel appeared first on Towards Data Science. Read More
