Adaptive multi-task positive-unlabeled learning for joint prediction of multiple chronic diseases using online shopping behaviors

Digitalization of daily life calls for new insights into preventing chronic diseases planted by long exposure to unhealthy lifestyles. This study explores an innovative idea of jointly predicting multiple chronic diseases using online shopping behaviors, targeted for times when e-commerce user experience has been assimilated into most people's everyday lives. We manually construct a novel dataset, in which each instance contains an online shopper's purchase records and two heuristic labels indicating if he/she is suffering from depression and diabetes, two lifestyle-related chronic diseases, respectively. As determining negative instances of chronic diseases among online shoppers is conceptually intractable, this new dataset consists of only 2219 depression and 2800 diabetes positive instances as well as 100000 unlabeled ones, which triggers an algorithmic challenge - multi-task positive-unlabeled (PU) learning. For this problem, an adaptive representation learning model is developed to enable generalizable features of online shopping behaviors to be shared across multiple tasks of chronic disease prediction by a task-specific attention mechanism. Our motivation is to leverage multimorbidity potentials, such as the bi-directional association between depression and diabetes, to make full use of the limited positive instances to improve predictive accuracy for each chronic disease simultaneously. In addition, a well-established loss function for PU learning is extended from single-task to multi-task to address the lack of negative instances during model training. Extensive experiments validate the effectiveness of the proposed approach and demonstrate its richness in interpreting buying patterns of online shoppers with chronic diseases.

Automated summary

This study presents an innovative approach to predict multiple chronic diseases using online shopping behaviors, leveraging an adaptive representation learning model and task-specific attention mechanism for feature sharing across tasks. The model addresses the lack of negative instances during training, demonstrating effectiveness in interpreting buying patterns of online shoppers with chronic diseases.