To assess students' programming skills, teachers spend considerable time to check the source codes submitted from students as a part of programming class. To reduce the burden, teachers may use some kinds of source code similarity measuring tools automate the task. However, it is difficult to differentiate a pair of source codes produced as assignments of programming class because the source codes are relatively short in length, simple in its structure, and similar each other in algorithm. In this study, we introduce FRef [1], the similarity measuring method that utilize the feature vectors calculated by using some reference data. The method showed effectiveness in searching similar source codes. We apply FRef in combination with a simple similarity measure based on software metrics aiming to achieve an auto-scoring function of the education support system for programming class in universities. We also discuss about reasonable criteria of correctness for auto-scoring function.