Feature based representation is inadequate in modeling different writing styles, irregularity in size, complicated structural relationships, cursivenes present in unconstrained handwritten devanagiri numerals. In this paper, this insufficiency is eliminated by giving graph based representation. Moreover, graphs are robust to similarity deformations as well. Graph Dissimilarity Space Embedding is explored to extract features from numeral graphs. The feature vector generated was trained on SVM with RBF Kernel. Efficacy of the method is corroborated by carrying extensive experiments on benchmark dataset. From this study graph based representation seems to be robust and powerful in representing handwritten devanagiri numerals.