Average Localised Proximity: A new data descriptor with good default one-class classification performance

One-class classification is a challenging subfield of machine learning in which so-called data descriptors are used to predict membership class based solely on positive examples that class, and no counter-examples. A number have been shown perform well previous studies one-class classification, like the Support Vector Machine (SVM), require setting one or more hyperparameters. There has systematic attempt date determine optimal default values for these hyperparameters, limits their ease use, especially comparison with hyperparameter-free proposals Isolation Forest (IF). We address this issue by determining hyperparameter across collection 246 problems derived from 50 different real-world datasets. In addition, we propose new descriptor, Average Localised Proximity (ALP) certain issues existing approaches nearest neighbour distances. Finally, evaluate performance using leave-one-dataset-out procedure, find strong evidence ALP outperforms IF other descriptors, as weak it SVM, making good choice.