WELLNESS PROFILING ON SOCIAL NETWORKS

Abstract

The increasing popularity of social media has encouraged health consumers to share, explore, and validate health and wellness information on social networks, which provide a rich repository of Patient Generated Wellness Data (PGWD). While data-driven healthcare has attracted a lot of attention from academia and industry for improving care delivery through personalized healthcare, limited research has been done on harvesting and utilizing PGWD available on social networks. This thesis focuses on learning wellness profiles of users, both at micro-level of individuals and macro-level of communities. Towards this end, we propose a unified framework and algorithms to perform the following tasks.

(1) To extract the wellness information of users, we propose a learning framework that utilizes the content information of microblogging messages as well as the relations among event categories to categorize messages into a wellness taxonomy.

(2) To learn the latent profile of users, we propose an approach which directly learns the embedding from longitudinal data of users, instead of vector-based representation. In particular, the proposed framework simultaneously learns a low-dimensional latent space as well as the temporal evolution of users in the wellness space. To construct an effective framework, we incorporate two types of wellness prior knowledge: (a) temporal progression of wellness attributes; and (b) heterogeneity of wellness attributes in the patient population. The proposed approach scales well to large datasets using parallel stochastic gradient descent.

(3) To learn the profile of user groups, we first integrate different social views of the network into a low-dimensional latent space representing users' profiles. We then learn the optimal community structure by imposing a similarity constraint over the affiliation vectors of the users, which seeks dense clusters of users in the latent space. We seamlessly incorporate prior knowledge about the community structure into the community discovery process and turn the process into an optimization problem, where community profile is constructed using a linear pooling operator integrating the profiles of the members.

To evaluate the effectiveness of the proposed framework, two large scale datasets were constructed by crawling social activities of diabetes patients in Twitter. Extensive experiments have demonstrated: (1) the importance of modeling both content information and events relation in wellness event extraction; (2) the significance of joint modeling temporality of wellness features and heterogeneity of the user in wellness profiling; (3) the importance of fusing all social behaviors for community discovery and profiling.