RUA. Repositorio Institucional de la Universidad de Alicante
Publication
Otros (Other)
Multivariate Automatic Tuning of Isolation Forest for Anomaly Detection in Critical Infrastructures: A Solution for Intelligent Information Systems
The Isolation Forest (IF) algorithm is effective in detecting anomalies in critical infrastructure, but its performance depends on the proper setting of five hyperparameters: sample size, number of trees, maximum tree depth, maximum number of features and detection threshold. Static tuning of these parameters is inefficient and poorly adaptable to dynamic environments. This paper proposes a multivariate autotuning method that automatically optimises these hyperparameters by: (1) adaptive adjustment of the sample size based on the standard deviation of the anomaly scores, (2) selection of the number of trees according to F1-score stabilisation, (3) control of the maximum depth based on the average isolation rate, (4) adjustment of the maximum number of features according to the variance of the data, and (5) optimisation of the detection threshold by minimisation of a cost function. The auto-tuning procedure has been validated in the detection of anomalies in drinking water networks, s howing an F1-score improvement of 7.5% and a reduction of the execution time by 22.55% compared to static configurations, demonstrating its feasibility for real-time systems.