Nearest Neighbor Algorithm for Positive and Unlabeled Learning with Uncertainty

This paper studies the problem of uncertain data classification under positive and unlabeled (PU) learning scenario. It proposes a novel algorithm, NNPU (nearest neighbor algorithm for positive and unlabeled learning), to handle this problem with two varieties, NNPUa and NNPUu. Experimental results on benchmark UCI datasets show that NNPUu, which considers the whole uncertain information on the datasets, has a better ability to classify unseen examples than NNPUa that considers the average value of uncertainty only. Furthermore, NNPU outperforms some existing algorithms such as NN-d, OCC (one-class classifier) and aPUNB in handling precise data.