Incomplete Mixed Data Outlier Detection based on Local Difference Information

Outlier detection is a fundamental task in data mining and knowledge discovery. For graph-based unsupervised outlier detection, it is typical to discover normal object behavior by utilizing the similarity neighborhood relationship among objects to identify outliers. However, the core challenge in outlier detection lies in identifying anomalies, errors, or rare instances that deviate from ordinary patterns. Directly leveraging object dissimilarity for outlier detection may be more straightforward and efficient. Accordingly, we propose a novel and efficient DRS (Distant Relative Stranger) model based on dissimilarity to discover outliers. Specifically, this study constructs a distant neighbor information network for incomplete mixed data based on object dissimilarity and uncertainty. By utilizing the local information of probabilistic graph models, the model characterizes the degree of anomaly for each object, achieving strong adaptability and high-performance unsupervised outlier detection. Through performance evaluation on 16 real-world datasets and comparison with 12 other models, it is evident that the proposed DRS model demonstrates the advantages of simplicity and stable performance in unsupervised outlier detection tasks. This research provides a new avenue and applicable approach for graph-based outlier detection studies.