With the recent surge of interest in machine learning, Positive and Unlabeled learning (PU learning) has also attracted much attention of scholars. A key bottleneck for addressing PU classification is the absence of training negative data, and thus many popular approaches belonging to the “two-step” strategy have been proposed. However, almost none of the existing two-step methods can thoroughly learn the feature information of samples, which makes the extracted negative samples unreliable and easily leads to undesirable results. Therefore, in this paper, we propose a two-phase projective dictionary pair learning (TPDPL) method for PU learning. The first phase of TPDPL determines reliable negatives by exploiting the reconstruction residuals and the second phase trains the DPL-based classifier with the extracted reliable negative and original positive samples to perform classification. Our experimental results demonstrate that the TPDPL approach can achieve highly competitive classification performance when compared with conventional and state-of-the-art PU learning algorithms. More importantly, due to the special dictionary pair learning framework, the computational complexity of TPDPL is extraordinarily low.