Analysis of Graph Construction Methods in Supervised Data Classification

Abstract

Graph-based methods have attracted a lot of attention in recent years, especially due to its inherent ability to capture properties of the networked data (e.g., structural and dynamical). Clustering, semi-supervised label propagation and, more recently, data classification are examples of tasks in which graph-based learning methods have obtained relevant results. In any of these tasks, the common approach is (i) to transform the feature vector data in a graph and then (ii) exploit some property uncovered by the network structure. However, most works have focused on the development of models to exploit the graph, while the graph construction step has been little explored. In this article, we conduct a preliminary study to evaluate supervised graph construction methods based on k-nearest neighbors (kNN) and ϵ-radius neighborhood (ϵN) criteria by employing a recently proposed classification technique based on the importance concept of complex networks. Experiments were conducted on artificial and real-world data sets, including the problem of invariant pattern recognition in images. The results show that the graph construction methods under study are able to deal with different configuration of problems (e.g., domain, features, etc). They also suggest that the combination between selective kNN and ϵN is more suitable in data sets with low level of mixture among the classes, while kNN seems slightly better in problems with higher noise levels.