Problem solving in physics and mathematics has been characterized in terms of five phases by Schonfeld and these have previously been used to describe also online and blended behavior. We argue that expanding the use of server logs to make detailed categorizations of student actions can help increase knowledge about how students solve problems.
We present a novel approach for analyzing server logs that relies on network analysis and principal component analysis. We use the approach to analyze student interactions with an online textbook that features physics problems. We find five 'components of behavioral structure': Complexity, Linear Length, Navigation, Mutuality, and Erraticism. Further, we find that problem solving sessions can be divided into three over-arching groups that differ in their Complexity and further into ten clusters that also differ on the other components. Analyzing typical sessions in each cluster, we find ten different behavioral structures, which we describe in terms of Schonfeld's phases. We suggest that further research integrates this approach with other methodological approaches to get a fuller picture of how learning strategies are employed by students in settings with online features.