Anomaly Detection: an Isolation Forest built from scratch. This project implements the Isolation Forest algorithm (Liu, Ting & Zhou, 2008) in pure Python and NumPy, without scikit-learn. It detects anomalies in numerical data.

How it works. Anomalies are rare and far from the rest of the data, so random splits isolate them quickly. The forest builds many random trees. The fewer splits it takes to isolate a point, the more likely that point is an anomaly.

What's in isolation_forest.py:

• c_factor(n): the normalisation factor. It's the average path length of an unsuccessful search in a binary search tree of n points, using the Euler–Mascheroni constant.
• anomaly_score(): the standard score s(x, n) = 2^(-E(h(x)) / c(n)). A score close to 1 means an anomaly; a score well below 0.5 means a normal point.
• Node: one node of a tree, either an internal split or a leaf.
• IsolationTree: builds one tree recursively. At each node it picks a random feature and a random split value between that feature's min and max. path_length() measures how deep a point goes before reaching a leaf.
• IsolationForest: the main model, with a scikit-learn-style API:
  • fit(X) trains n_estimators trees, each on a random subsample of max_samples points, with depth capped at log2(sample_size). It then sets the decision threshold from the contamination rate.
  • decision_function(X) returns an anomaly score for each point.
  • predict(X) returns -1 for an anomaly and 1 for a normal point.
• Built-in demo: 200 normal points clustered around (0, 0), plus 10 outliers spread across [-4, 4]. It trains the forest with 5% contamination and prints the score and label for each outlier.

Stack: Python 3.14 and NumPy 2.5.3, in a local .venv.

A bug to fix: decision_function normalises with min(X.shape[0], self.max_samples), the size of the data being scored, not the subsample size used in training. The demo scores only 10 outliers, so it normalises with c(10) instead of c(210). Those scores are then on a different scale from the threshold, which was computed on all 210 training points. The fix is to save the training sample_size in fit() (for example as self.sample_size_) and reuse it in decision_function.

I can also put this description into your readme.txt, in English or in French to match the code comments.