Using Unsupervised Learning for Graph Construction in Semi-supervised Learning with Graphs

Semi-supervised Learning with Graphs can achieve good results in classification tasks even in difficult conditions. Unfortunately, it can be slow and use a lot of memory. The first important step of the graph-based semi-supervised learning approaches is the construction of the graph from the data, where each data-point usually becomes a vertex in the graph - a potential problem with large amounts of data. In this paper, we present a graph construction method that uses an unsupervised neural network called growing neural gas (GNG). The GNG instance presents a intelligent stopping criteria that determines when the final network configuration maps correctly the input-data points. With that in mind, we use the final trained network as a reduced input graph for the semi-supervised classification algorithm, associating original data-points to the neurons they have activated in the unsupervised training process.