Abstract
Detecting railroad station anomalies is a critical task prior to segmentation and making optimization decisions for each cluster. Three types of anomalies (local clustered, axis paralleled, and surrounded by normal instances) caused by the specialty of railroad operations bring the existing methods non-trivial challenges in detecting them accurately and efficiently. To tackle this limitation of existing methods, this paper proposes a novel anomaly detection method named Huffman Anomaly Detection Forest (HuffForest) to detect station anomalies, which leverages Huffman encoding to measure abnormalities in certain railroad scenarios with high accuracy. The proposed method establishes a Huffman forest by constructing trees from the perspective of data points and subsequently computes anomaly scores of instances considering both local and global information. A sampling-based version is also developed to improve scalability for large datasets. Taking advantage of the encoding mechanism, the proposed method can effectively recognize the underlying patterns of railroad stations and detect outliers in various complicated scenarios where the conventional methods are not reliable. Experiment results on both synthesized and public benchmarks are demonstrated to show the advances of the proposed method compared to the state-of-the-art isolation forest (iForest) and local outlier factor (LOF) methods on detection accuracy with an acceptable computational complexity.