Little is known about the ways that underrepresented students in online STEM courses interact and behave differently from their peers, or whether online courses offer learning opportunities that can better suit these under-served populations. The current study examines the logged behavioral patterns of 470 university students, spanning 3 years, who were enrolled in an online introductory STEM course. Cross-validated data mining methods were applied to their interaction logs to determine if first generation, non-white, female, or non-Traditional (≥ 23 years old) students could be classified by their behaviors. Model classification accuracies were evaluated with the Matthews Correlation Coefficient (MCC). First generation (MCC = .123), non-white (MCC = .153), female (MCC = .183) and non-Traditional students (MCC = .109) were classified at levels significantly above chance (MCC = 0). Follow-up analyses of predictive features showed that first-generation students made more quiz attempts, non-white students interacted more during night hours (8pm-8am), female students submitted quizzes earlier, and non-Traditional students accessed discussion forums less than their peers. We show that understanding behaviors is crucial in this context because behaviors in the first two weeks alone (e.g., discussion forumparticipation, number of logins) predicted eventual grade in the course (MCC = .200). Implications are discussed, including suggestions for future research as well as interventions and course features that can support underrepresented STEM students in online learning spaces.