Improved Isolation Forest Algorithm Based on Feature Weights and Local Data Relationships
Haoyuan Li, Yu Jiang · 2023
Aiming at the problem of declining accuracy of isolation forest algorithm in high-dimensional data and misclassification caused by the considering only global data relations, this paper proposes an improved isolation forest algorithm based on feature weight and local data (FLiForest). Firstly, the method calculates the mutual information of feature and density for subsamples, and sets the feature weights according to the mutual information. Secondly, the nearest neighbor method is used to find the anomaly score of the local neighborhood of the test object, and the local score matrix is formed. Finally, the abnormal scores in the local score matrix are combined to determine the abnormal scores of the test data. Based on eight public anomaly detection data sets, the proposed method is compared with eight advanced anomaly detection algorithms. The area under the receiver operating characteristic curve (AUC) and the average accuracy (AP) were higher by 8%-30% and 18%-43% on average.