Automatically Design Distance Functions for Graph-Based Semi-Supervised Learning

Abstract

Automatic data classification is often performed by supervised learning algorithms, producing a model to classify new instances. Reflecting that labeled instances are expensive, semisupervised learning (SSL) methods prove to be an alternative to performing data classification, once the learning demands only a few labeled instances. There are many SSL algorithms, and graph-based ones have significant features. In particular, graph-based models grant to identify classes of different distributions without prior knowledge of statistical model parameters. However, a drawback that might influence their classification performance relays on the construction of the graph, which requires the measurement of distances (or similarities) between instances. Since a particular distance function can enhance the performance for some data sets and decrease to others, here, we introduce a novel approach, called GEAD, a Grammatical Evolution for Automatically designing Distance functions for Graph-based semi-supervised learning. We perform extensive experiments with 100 public data sets to assess the performance of our approach, and we compare it with traditional distance functions in the literature. Results show that GEAD is capable of designing distance functions that significantly outperform the baseline manually-designed ones regarding different predictive measures, such as Micro-F1, and Macro-F1.