Substructure counting graph kernels for machine learning from RDF data

Highlights

• Systematic graph kernel framework for RDF.

• Fast computation algorithms.

• Low frequency labels and hub removal on RDF to enhance machine learning.

Abstract

In this paper we introduce a framework for learning from RDF data using graph kernels that count substructures in RDF graphs, which systematically covers most of the existing kernels previously defined and provides a number of new variants. Our definitions include fast kernel variants that are computed directly on the RDF graph. To improve the performance of these kernels we detail two strategies. The first strategy involves ignoring the vertex labels that have a low frequency among the instances. Our second strategy is to remove hubs to simplify the RDF graphs. We test our kernels in a number of classification experiments with real-world RDF datasets. Overall the kernels that count subtrees show the best performance. However, they are closely followed by simple bag of labels baseline kernels. The direct kernels substantially decrease computation time, while keeping performance the same. For the walks counting kernel this decrease in computation time is so large that it thereby becomes a computationally viable kernel to use. Ignoring low frequency labels improves the performance for all datasets. The hub removal algorithm increases performance on two out of three of our smaller datasets, but has little impact when used on our larger datasets.

Introduction

In recent years graph kernels have been introduced as a promising method to perform data mining and machine learning on Linked Data and the Semantic Web. These methods take Resource Description Framework (RDF) data as input.

One main advantage of this approach is that the techniques are therefore very widely applicable to all kinds of Linked Data. Almost no assumptions are made on the semantics of the data and its content, other than that it is in RDF. So, additionally, these methods require very little knowledge of Semantic Web technologies to be employed.

Another advantage is the host of existing machine learning algorithms, called kernel methods  [1], [2], that can be used with these graph kernels. The most well known of these algorithms is the Support Vector Machine (SVM) for classification and regression. However, algorithms exist for ranking  [3], [4], clustering  [5], outlier detection  [6], etc., which can all be used directly with these kernels. More recently, interest has increased for large scale linear classification  [7] for larger datasets, for which a number of graph kernels can also be used.

In this paper we give a comprehensive overview of graph kernels for learning from RDF data. We introduce a framework for these kernels, which are based on counting different graph substructures, that encompasses most of the graph kernels previously introduced for RDF, but also introduces new variants. The framework includes fast kernel variants that are computed directly on the RDF graph. We also detail the necessary adaptation of the Weisfeiler–Lehman graph kernel  [8] needed to compute a number of kernels in our framework. Furthermore, we give two strategies to further improve the machine learning performance with these kernels. The first strategy ignores vertex labels which have a low frequency of occurrence among the instances and the second strategy removes hubs to simplify the RDF graphs. All of our kernels defined in the framework can be used with large scale linear classification methods.

The kernels are studied in a number of classification experiments on different RDF datasets. The goal of these experiments is to study the influence of the different choices for graph kernels defined in our framework. It turns out that kernel performance differs per dataset. Overall, kernels that count subtrees in the graphs are the best choice. However, simple bag of labels baseline kernels also perform well and are significantly cheaper to compute. The strategy to ignore low frequency labels has a positive effect on performance in all tasks, whereas the hub removal strategy only has a positive effect in a number of tasks and has no influence for larger datasets.

The work presented in this paper consolidates and expands our earlier papers on graph kernels for RDF  [9], [10] and hub removal [11].

The rest of this paper is structured as follows. We begin with an overview of related work. In Section  3 we introduce our kernel framework and algorithms. Section  4 covers our experiments with these kernels. We end with conclusions and suggestions for future work.