Streaming Semi-Supervised Adaptive Isolation Forest
Abstract
Anomaly detection in data streams is vital in areas such as network security, finance, and IoT, but it is hindered by the scarcity and cost of labels. We propose the Streaming Semi-Supervised Adaptive Isolation Forest (S3AIF), which extends the Adaptive Isolation Forest to partially labelled data, evolving streams by using discrimination power as a unifying principle to: (i) weight isolation trees when scoring; (ii) replace the least discriminative tree at each window; and (iii) select query instances for active learning by path-length disagreement. Across 22 benchmark datasets with a labelling budget below 1%, S3 AIF outperforms HiLAD in both ROC-AUC and PR-AUC while running more than three times faster and ranks first against 10 unsupervised baselines in ROC-AUC. An ablation confirms that discrimination-power-based tree replacement is the main driver of S3 AIF's performance.