CRAFTIIF: Cross-Resolution Analytic Four-Type Interpretable Isolation Forest for Multivariate Time Series Anomaly Detection

Problem
Anomaly detection in multivariate time series data is often limited by existing methods that focus on only one or two types of anomalies, which include point, distributional, temporal, and collective anomalies. This paper addresses the gap in capability for a comprehensive, unsupervised approach that can detect all four types without requiring dataset-specific tuning. The work is presented as a preprint, indicating it has not yet undergone peer review.

Method
The authors propose CRAFTIIF (Cross-Resolution Analytic Four-Type Interpretable Isolation Forest), which employs a fully unsupervised framework. CRAFTIIF generates K=500 random analytic wavelet feature draws from four families: Morlet, Difference of Gaussian (DOG), Haar, and Coiflet, each tailored to capture specific anomaly characteristics. The framework utilizes five structured Isolation Forests (IFs)—one for each anomaly type and a meta-IF for compound anomalies. An adaptive Otsu/MAD thresholding mechanism is implemented to automatically calibrate detection across a wide range of anomaly rates (0.1% to 69.2%). The design allows for direct anomaly-type attribution through branch firing, eliminating the need for post-hoc explanations.

Results
CRAFTIIF was evaluated on all 19 datasets from the mTSBench benchmark, achieving a mean F1 score of 0.228 across all datasets and an F1 score of 0.322 on the 13 datasets where anomalies were detectable. It ranked first among 25 evaluated methods on the Variable Uncertainty Score-Precision-Recall (VUS-PR) metric, with a score of 0.463, surpassing the previous best score of 0.329 by 40.7%. The authors also established a diagnostic framework that includes oracle F1 scores, detectability limits, and branch separation ratios, identifying 6 out of 19 datasets as fundamentally undetectable by any unsupervised method. Ablation studies demonstrated the importance of adaptive thresholding (+38% F1), the four-branch structure (+20%), and the meta-IF (+23%) for performance enhancement.

Limitations
The authors acknowledge that their method may not be effective on datasets that are inherently undetectable by any unsupervised approach, as indicated by their diagnostic framework. Additionally, the reliance on wavelet feature extraction may limit performance in scenarios where other feature representations could be more effective. The paper does not address potential computational overhead associated with generating a large number of features or the scalability of the method to larger datasets.

Why it matters
CRAFTIIF’s ability to detect multiple types of anomalies in a fully unsupervised manner with high interpretability has significant implications for real-world applications in fields such as finance, healthcare, and industrial monitoring. The framework’s design allows for direct attribution of detected anomalies, which can enhance decision-making processes. This work contributes to the ongoing discourse in anomaly detection methodologies, as discussed in related literature, and is available on arXiv.