A Generative Model for Discovering Action-Based Roles and Community Role Compositions on Community Question Answering Platforms
This paper proposes a generative model for discovering user roles and community role compositions in Community Question Answering (CQA) platforms. While past research shows that participants play different roles in online communities, automatically discovering these roles and providing a summary of user behavior that is readily interpretable remains an important challenge. Furthermore, there has been relatively little insight into the distribution of these roles between communities. Does a community’s composition over user roles vary as a function of topic? How does it relate to the health of the underlying community? Does role composition evolve over time? The generative model proposed in this paper, the mixture of Dirichlet-multinomial mixtures (MDMM) behavior model can (1) automatically discover interpetable user roles (as probability distributions over atomic actions) directly from log data, and (2) uncover community-level role compositions to facilitate such cross-community studies.
A comprehensive experiment on all 161 non-meta communities on the StackExchange CQA platform demonstrates that our model can be useful for a wide variety of behavioral studies, and we highlight three empirical insights. First, we show interesting distinctions in question-asking behavior on StackExchange (where two distinct types of askers can be identified) and answering behavior (where two distinct roles surrounding answers emerge). Second, we find statistically significant differences in behavior compositions across topical groups of communities on StackExchange, and that those groups that have statistically significant differences in health metrics also have statistically significant differences in behavior compositions, suggesting a relationship between behavior composition and health. Finally, we show that the MDMM behavior model can be used to demonstrate similar but distinct evolutionary patterns between topical groups.