A Study of Transductive Graph-Based Regression

Regression methods play an important role in many real-world applications such as econometric, pattern recognition, and prediction of protein chains to cite a few tasks. In the semi-supervised transductive learning problem, some studies present graph-based regression solutions. The spectral network is the main topology adopted in these solutions. This network is characterized by being a fully connected network. However, in the semi-supervised transductive learning environment where few elements are valued and therefore little data will be used in a linear regression operation, therefore there would not be a need for many connections in this situation. A network with fewer connections accompanied by a technique based on continuous value propagation graphs may be superior to solutions found in the literature. Thus, we proposed combining two topological network models to benefit the predicted values in the semi-supervised context. First, we combine two well-known network construction models, namely, spectral graph and k-nearest neighbors (k-NN), to ensure a hybrid-connected network. Second, as the inference strategy, we use a network technique combining Random Walks and the regression task, which turns forward the propagation of continuous values by the network in unlabeled objects using the least-squares regression method. An empirical analysis of different datasets shows that our strategy surpasses traditional approaches considering the measures used in the evaluation.