An enhanced anomalies detection method based on isolation forest and fuzzy set

Most existing anomaly detection models are model-based approaches to generate the pattern of normal instances, then mark instances that do not conform to the normal pattern as anomalies. This kind of algorithms are often complex. Isolation Forest, on the contrary, isolates anomalies rather than generating patterns of normal instances, it has linear time complexity and low memory requirements. However, the process of iForest sampling to construct Isolated Trees is random, and the algorithm preformance is not stable. In this paper, an enhanced anomalies detection method based on iForest and fuzzy set is proposed (short for, FForest). In the sampling process, the subset of possible anomalies is obtained by calculating the quartile points in advance instead of random sampling. At the same time, the fuzzy membership degree is used to measure the anomalyscore, so as to enhance the interpretable of the algorithm. The experiments with seven real-world data sets demonstrate our method outperforms four baseline methods.