Despite the last 40 years of research showing that computer-aided diagramming tools improve student learning, very little research reveals the cognitive processes that explain why and how diagramming tools and specific features of the tools affect learning. This study developed a tool that graduate students used to diagram and analyze arguments as the tool mined each students' actions. The mined data was used to develop algorithms to operationalize and measure the use of backward, forward, breadth, depth-first reasoning. Regression models were then compared to identify which algorithm produced measures that best predicted diagram scores, and to determine the combined and relative impact of each reasoning process on diagram scores. The findings show that observing the placement and location of the first five nodes moved and positioned on screen in relation to the location of the previously moved node provides sufficient data to generate backward/forward and breadth/depth-first ratio scores that predict diagram scores, while individual frequency counts of each of the four processes do not predict scores. The best-fit regression model using the ratio scores show that students using more backward and depth-first processing construct diagrams with higher scores -- scores based on criteria that gauge the depth of analysis and not just the number of correct diagram links. This study presents new tools, methods, and new lines of inquiry to advance research on ways to integrate learning analytics into computer-aided diagramming tools.