Towards positive unlabeled learning for parallel data mining: a random forest framework

Parallel computing techniques can greatly facilitate traditional data mining algorithms to efficiently tackle learning tasks that are characterized by high computational complexity and huge amounts of data, to meet the requirement of real-world applications. However, most of these techniques require fully labeled training sets, which is a challenging requirement to meet. In order to address this problem, we investigate widely used Positive and Unlabeled (PU) learning algorithms including PU information gain and a newly developed PU Gini index combining with popular parallel computing framework - Random Forest (RF), thereby enabling parallel data mining to learn from only positive and unlabeled samples. The proposed framework, termed PURF (Positive Unlabeled Random Forest), is able to learn from positive and unlabeled instances and achieve comparable classifcation performance with RF trained by fully labeled data through parallel computing according to experiments on both synthetic and real-world UCI datasets. PURF is a promising framework that facilitates PU learning in parallel data mining and is anticipated to be useful framework in many real-world parallel computing applications with huge amounts of unlabeled data.