Internal Evaluation of Unsupervised Outlier Detection

摘要: Although there is a large and growing literature that tackles the unsupervised outlier detection problem, evaluation of results still virtually untouched in literature. The so-called internal evaluation, based solely on data assessed solutions themselves, required if one wants to statistically validate (in absolute terms) or just compare relative provided by different algorithms parameterizations given algorithm absence labeled data. However, contrast cluster analysis, where indexes for validation clustering have been conceived shown be very useful, domain, this problem has notably overlooked. Here we discuss provide solution results. Specifically, describe an index called Internal, Relative Evaluation Outlier Solutions (IREOS) can evaluate candidate solutions. Initially, designed binary only, referred as top-n We then extend IREOS general case non-binary solutions, consisting scorings. also adjust chance extensively it several experiments involving collections synthetic real datasets.