Knowledge components (KCs) define the underlying skill model of intelligent educational software, and they are critical to understanding and improving the efficacy of learning technology. In this research, we show how learning curve analysis is used to fit a KC model--one that was created after use of the learning technology--which can then be improved by human-centered data science methods. We analyzed data from 417 middle-school students who used a digital learning game to learn decimal numbers and decimal operations. Our initial results showed that problem types (e.g., ordering decimals, adding decimals) capture students' performance better than underlying decimal misconceptions (e.g., longer decimals are larger). Through a process of KC model refinement and domain knowledge interpretation, we were able to identify the difficulties that students faced in learning decimals. Based on this result, we present an instructional redesign proposal for our digital learning game and outline a framework for post-hoc KC modeling in a tutoring system. More generally, the method we used in this work can help guide changes to the type, content and order of problems in educational software. [For the full proceedings, see ED599096.]