Health Consulting Services Recommendation Considering Patients’ Decision-Making Behaviors: A CNN and Multiarmed Bandit Approach

For online healthcare community (OHC) platforms, recommending suitable health consulting services (HCSs) to patients is a complex engineering decision-making problem. Many HCSs recommendation methods have been proposed; however, they encounter the serious cold start problem and overlook the patients' decision preferences. To address these issues, this paper proposes a hybrid recommendation method that considers both the patients' disease features and decision-making behaviors through a convolutional neural network and multi-armed bandit approach (CNN-MAB). The proposed CNN-MAB method includes three modules: (1) a CNN-based feature learning module to extract latent features of patients and HCSs, (2) a similarity analysis module to generate initial recommendations by comparing new and historical patients and HCSs, and (3) a contextual MAB-based decision-making behavior learning module to refine the recommendations based on patient preferences. Experiments conducted on an online diabetes community demonstrate that the proposed CNN-MAB method outperformed several benchmark methods by 22.8%, 17.1%, and 16.4% in terms of recall, MRR, and coverage, respectively. We found that new patients' perceptions of the historical patients' disease features proved to be more influential in HCS selection than decision preferences. In addition, the dynamic MAB method demonstrates superior effectiveness in analyzing decision behaviors compared with conventional questionnaire-based or review-based approaches.