Considering each of the visual features as one modality in image annotation task, efficient fusion of different modalities is essential in graph-based learning. Traditional graph-based methods consider one node for each image and combine its visual features into a single descriptor before constructing the graph. In this paper, we propose an approach that constructs a subgraph for each modality in such a way that edges of subgraph are determined using a search-based approach that handles class-imbalance challenge in the annotation datasets. Multiple subgraphs are then connected to each other to have a supergraph. This follows by introducing a learning framework to infer the tags of unannotated images on the supergraph. The proposed approach takes advantages of graph-based semi-supervised learning and multi-modal representation simultaneously. We evaluate the performance of the proposed approach on different datasets. The results reveal that the proposed approach improves the accuracy of annotation systems.