Supplementary Material for
On the Evaluation of Unsupervised Outlier Detection: Measures, Datasets, and an Empirical Study
by G. O. Campos, A. Zimek, J. Sander, R. J. G. B. Campello, B. Micenková, E. Schubert, I. Assent and M. E. Houle
Data Mining and Knowledge Discovery 30(4): 891-927, 2016, DOI: 10.1007/s10618-015-0444-8


The 10 classes contained in this data set correspond to the digits from 0 to 9, with examples created by different hand writings. Class 4, defined here as outlier, was downsampled to only 20 objects. After the preprocessing, this database has 16 numeric attributes and 9868 instances, divided into 20 outliers (0.2%) and 9848 inliers (99.8%). This dataset is already normalized, i.e., all 16 attributes (spatial coordinates) have the same range [0,100]. It has been used in this form in [1,2].


[1] H.-P. Kriegel, P. Kroeger, E. Schubert, and A. Zimek. Interpreting and unifying outlier scores. In Proc. SDM, pages 13-24, 2011.
[2] E. Schubert, R. Wojdanowski, A. Zimek, and H.-P. Kriegel. On evaluation of outlier rankings and outlier scores. In Proc. SDM, pages 1047-1058, 2012.

Download all data set variants (2.1 MB). Access original data (merge train and test [pendigits.tes and pendigits.tra])