This paper, entitled Graph-based Bag-of-Words for Classification, introduces the Bag of Graphs (BoG), a Bag-of-Words model that encodes in graphs the local structures of a digital object. It presents a formal definition, introducing concepts and rules that make this model flexible and adaptable for different applications. It is defined two BoG-based methods – Bag of Singleton Graphs (BoSG) and Bag of Visual Graphs (BoVG), which create vector representations for graphs and images, respectively. The hypothesis explored in this paper is that the combination of graphs with the BoW model can create a discriminant and efficient representation based on local structures of an object, leading to fast and accurate results in classification tasks. The rationale is that the two representations are complementary and can help each other overcome their individual deficiencies.
The authors evaluate the Bag of Singleton Graphs (BoSG) for graph classification on four datasets of the IAM repository, obtaining significant results in accuracy and execution time. The method Bag of Visual Graphs (BoVG) is evaluated for image classification on Caltech and ALOI datasets, and for remote sensing image classification on images of Monte Santo and Campinas datasets. This framework opens possibilities for retrieval, classification, and clustering tasks on large datasets that use graph-based representations impractical before due to the complexity of inexact graph matching.
Fernanda B. Silva, Rafael de O. Werneck, Siome Goldenstein, Salvatore Tabbone, Ricardo da S. Torres, Graph-based Bag-of-Words for Classification, Pattern Recognition (2017), doi: 10.1016/j.patcog.2017.09.018