Semi-Supervised Regression Algorithm Based on Optimal Combined Graph

In order to construct a high-quality graph to improve the learning accuracy, a new semi-supervised regression algorithm is proposed. According to all labeled and unlabeled samples, multiple graphs with different structures are constructed by using different edge-selection strategies and edge-measurement methods. Each graph corresponds to a basic graph kernel. Following that, a combined graph kernel is created by carrying out a convex optimization operation on these basic graph kernels. We can further obtain an optimal combined graph by calculating a pseudo-inverse of the combined graph kernel. Based on the optimal combined graph, a harmonic function is applied to solving the semi-supervised regression problem. Experimental results on typical artificial function and UCI real datasets show that, compared with other graph-based semi-supervised regression algorithms, the proposed algorithm has higher prediction accuracy even though its control parameters are not settled as optimum values.