Feature extraction is a crucial step in any computer aided diagnosis (CAD) system for melanoma diagnosis. Therefore, it is important to select features that are able to efficiently characterize the properties of the different types of lesions. Local features that separately characterize and distinguish different regions of the lesions have been shown to provide good descriptors for these skin lesions. Two powerful methods can be used to obtain local features: bag-of-features (BoF) and sparse coding (SC). Both methods have been applied to dermoscopy with promising results. However, a comparison between the two strategies is lacking. In this work, we fill this gap by developing a framework to compare the two methods in the melanoma diagnosis task. The results show that SC significantly outperforms BoF, achieving sensitivity = 85.5% and specificity = 75.1% versus sensitivity = 81.7% and specificity = 66.5%.