In sentiment classification, conventional supervised approaches heavily rely on a large amount of linguistic resources, which are costly to obtain for under-resourced languages. To overcome this scarce resource problem, there exist several methods that exploit graph-based semi-supervised learning (SSL). However, fundamental issues such as controlling label propagation, choosing the initial seeds, selecting edges have barely been studied. Our evaluation on three real datasets demonstrates that manipulating the label propagating behavior and choosing labeled seeds appropriately play a critical role in adopting graph-based SSL approaches for this task.