When data is expressed in a graph which consists of nodes and edges, there is an advantage that a relationship of data can be visually expressed. With a graph, machine learning can perform the prediction tasks. However, most of the real world data is sparse, and the number of labeled data is very small. In this case, machine learning cannot provide sufficient inference. In order to make prediction performance better in machine learning using graph, the graph should contain a lot of information (dense graph), or a lot of label information. To motivate this, we propose a graph enhancement learning. It consists of three algorithm; (1) edge complementation, (2) partial graph fusion, and (3) label transferring. Edge complementation is incremental learning to enhance disconnected edges in a sparse graph. Partial graph fusion is one of graph integration method. Combinations of many partial graphs help prediction tasks. To circumvent deficiency of lack of labeled data, we propose label transferring algorithms which generate pseudo-labels. The graph enhancement learning was applied to the biomedical graph. The edges of the disease network were complemented to find co-occurrence disease. In addition, the edge of the drug network was complemented to find the candidate drug. Gene networks combined with partial gene networks to identify key target genes for immune disease. To circumvent deficiency of lack of labeled data, we propose label transferring algorithms which generate pseudo-labels. And it is applied five benchmark dataset. When edges are not enough or labeled data is insufficient, the prediction performance of machine learning can be improved through graph enhancement learning.
