Info-Detection: An Information-Theoretic Approach To Detect Outlier

Outlier detection is one of major task in unsupervised learning. We propose a cluster analysis based outlier detection method called Info-Detection. Info-Detection determines the number of outliers automatically and captures the global property of the provided data. To implement Info-Detection and overcome the global computational complexity, we use principal sequence of partition, which we improve one order of magnitude faster than the original version. Experiments show that compared with other outlier detection methods, Info-Detection achieves better accuracy with an affordable time overhead.