A new supervised outlier detection method for hybrid data

Outlier detection is an important research topic in data mining. The hybrid data contains numerical features, categorical features and missing information values at the same time, providing a variety of attribute descriptions, which can adapt to more practical application scenarios. However, the existing methods for outlier detection based on rough set theory rely predominantly on unlabeled data, and there are few studies focusing on hybrid data. This article investigates a supervised outlier detection for hybrid data based on conditional information entropy. First, the distance between information values in a hybrid information system (HIS) is introduced, and a variable parameter \(\lambda\) $\lambda$ controlling the distance is provided. Then, conditional information entropy in a HIS is defined to measure the uncertainty of this HIS, and three metrics are constructed to reflect the abnormal degree of each object. Next, an outlier factor based on conditional information entropy is comprehensively established, and a corresponding algorithm (is called CIEOD) is designed. Finally, CIEOD is applied to 18 UCI data sets and compared with other eight outlier detection algorithms. The parameter \(\lambda\) $\lambda$ in CIEOD is analyzed. To evaluate the performance of these algorithms, two criteria are used: AUC value and F1-measure. The experimental results indicate that the designed algorithm has better effectiveness and superiority.