Many practical computing problems concern large graphs. Some examples include the Web graphs, various social networks and molecular datasets. The scale of these graphs introduces challenges to their ecient processing. One of the main issues in such problems is that most of the mentioned datasets cannot be t in the memory. In this paper, we present a new data fragment framework for graph mining. The original dataset is divided into a xed number of fragments, associated with the number of the graphs in each dataset. Then each fragment is mined individually using a well-known graph mining algorithm (i.e. FSG or gSpan) and the results are combined to generate global results. A major problem in fragmenting graphs is concerning on similarity or dissimilarity of them. Another problem corresponds to the completeness of the output which will be discussed in this paper.