Disease profiling in pharmaceutical E-commerce

Abstract Pharmaceutical e-commerce platforms need to know users’ disease for product recommendation. However, symptom-based disease prediction is not applicable in this context due to the lack of symptom data. In this study, utilizing only purchase data, a prudent and iterative naive Bayesian algorithm is proposed for disease prediction with the following advantages. First, it utilizes the freely available drug descriptions to identify some positive samples of a disease, hence converts the problem into a positive and unlabeled learning problem and avoids the costly training samples construction. Second, it proposes a prudent process to refine the quality of potentially positive samples in an iterative Bayesian learning process. This process also involves only one classifier while prior methods often involve multiple classifiers. This process enriches the scant strategies for positive sample selection in positive and unlabeled learning. The test on three liver diseases and three cardiovascular diseases indicate that the proposed algorithm could achieve a precision of 98.64% and recall of 90.90% across six diseases, which are superior to most of the benchmark algorithms.