We describe preliminary applications of network analysis techniques to eye-tracking data collected during a collaborative learning activity. This paper makes three contributions: first, we visualize collaborative eye-tracking data as networks, where the nodes of the graph represent fixations and edges represent saccades. We found that those representations can serve as starting points for formulating research questions and hypotheses about collaborative processes. Second, network metrics can be computed to interpret the properties of the graph and find proxies for the quality of students' collaboration. We found that different characteristics of our graphs correlated with different aspects of students' collaboration (for instance, the extent to which students reached consensus was associated with the average size of the "strongly connected components" of the graphs). Third, we used those characteristics to predict the quality of students' collaboration by feeding those features into a machine-learning algorithm. We found that among the eight dimensions of collaboration that we considered, we were able to roughly predict (using a median-split) students' quality of collaboration with an accuracy between ~85 and 100%. We conclude by discussing implications for developing "collaboration-sensing" tools, and comment on implementing this approach for formal learning environments.