Abstract: Graph based transductive classifiers are dependent on graph structure. Because of redundant and noisy features in high dimensional data, a graph, constructed from these data, can not reflect their distribution information faithfully. Consequently, the performance of a transductive classifier is downgraded. To address this problem, a multiple graphs construction scheme is introduced and applied into transductive classification. The scheme generates firstly several random subspaces and applies semi-supervised discriminative analysis in each subspace. Next, it trains a transductive classifier in each discriminative subspace. And finally, by voting rule, it fuses these classifiers as an ensemble classifier. Empirical results show that, in comparison with other transductive classifiers, the proposed ensemble classifier is more precise and robust to parameters selection.