An inductive semi-supervised learning approach for the Local and Global Consistency algorithm

Graph-based semi-supervised learning (SSL) algorithms learn through a weighted graph generated from both labeled and unlabeled examples. Despite the effectiveness of these methods on a variety of application domains, most of them are transductive in nature. Therefore, they are uncapable to provide generalization for the entire sample space. One of the most effective graph-based SSL algorithms is the Local and Global Consistency (LGC), which is formulated as a convex optimization problem that balances fitness on labeled examples and smoothness on the weighted graph through a Laplacian regularizer term. In this paper, we provide a novel inductive procedure for the LGC algorithm, called Inductive Local and Global Consistency (iLGC). Through experiments on inductive SSL using a variety of benchmark data sets, we show that our method is competitive with the commonly used Nadaraya-Watson kernel regression when applying the LGC algorithm as basis classifier.