Social Network Analysis to Delineate Interaction Patterns That Predict Weight Loss Performance

Abstract

Social media is an interesting, relatively new topic in health and self-management, which is generating enormous amounts of data, but little is yet known about its effect on the health of participants. The goal of this study is to determine online interaction behaviours that predict weight loss performance. The problem is modelled as a binomial classification task for predicting whether a patient would lose significant weight, based on analysis of two obesity online communities. An expansion-reduction method was developed for the patient feature vector, where the expansion is based on concatenating network structure features and the reduction is based on feature subset selection. Further, empirical evaluation of classifiers was done on the datasets, before and after the expansion. Based on feature subset selection, centrality measures such as degree and betweenness were more predictive than basic demographic features. Top performers, compared with bottom performers, were significantly more active online and connected to more than one sub-community (at 95% CI and p<;.05). In terms of classification, we found naive Bayes and decision tree methods had superior performance on the datasets, drastically reducing the false positive (FP) rate in some instances, and reaching a maximum F-score of 0.977, precision of 0.978 and AUC of 0.996. Current findings are consistent with previous reports that amount of online engagement correlates with weight loss, but our findings speak further to the types of engagement that yield best results.