Peer grading is an approach increasingly adopted for assessing students in massive on-line courses, especially for complex assignments where automatic assessment is impossible and the ability of tutors to evaluate and provide feedback at scale is limited. Unfortunately, as students may have different expertise, peer grading often does not deliver accurate results compared to human tutors. In this paper, we describe and compare different methods, based on graph mining techniques, aimed at mitigating this issue by combining peer grades on the basis of the detected expertise of the assessor students. The possibility to improve these results through optimized techniques for assessors' assignment is also discussed. Experimental results with both synthetic and real data are presented and show better performance of our methods in comparison to other existing approaches.