Relational Learning: A Web-Page Classification Viewpoint

This paper organises some general observations on Relational Learning which arose from research into classifying Web pages. The motivation for this piece is to contribute towards developing a broad overview of the field, so as to understand which aspects of Relational Learning are common to all domains, and which aspects are peculiar to specific domains. Hence the views presented here are necessarily only one piece of the puzzle, and it is hoped that analogous perspectives from other practitioners will improve upon the picture being developed here. With that in mind, let’s take a whirlwind tour of Relational Learning as seen through the eyes of someone interested in classifying Web pages. 1 Relational Learning Problems Relational learning problems, by definition, require data in some relational format. However, there is still some variety in the nature of the classification problems that are considered. Being aware of these variations can be useful when it comes to choosing a particular learning algorithm for the task at hand. The following general categories of relational learning problem are encountered when considering machine learning on the Web. Do problems from other domains fit into this categorisation? Are there more categories? 1.1 Classifying Graphs Instances Problems like classifying Web sites fall into this category. Instances are graphs, but usually not connected to each other and the task is to classify the full relational description of an instance. 1.2 Classifying Nodes Instances Problems like classifying Web pages fall into this category. Instances are nodes, and paths may or may not exist between them. The task is to classify an instance using both the properties of that instance and properties of the “neighbourhood” of that instance. Since nodes can appear in the neighbourhood of more than one instance, thorny independence issues can arise when trying to evaluate whether such shared neighbours are useful predictors or not. 1.3 Classifying Node Tuples Instances Problems like classifying pairs of Web pages (e.g. is this course being taught by this professor) fall into this category. Instances are tuples of nodes, normally connected in some way. The task is to classify an instance using information about how the nodes relate to each other. For the moment, I’ve not broken out edge classification into a separate section. Edge classification on the Web seems to be covered well enough by classifying the pair of web pages linked by the edge as a tuple. Are there compelling examples from other domains that warrant edge classification being considered independently? 2 Relational Learning Features Being aware of what kinds of regularity your particular relational learning problem may exhibit is important when designing and evaluating algorithms for that problem. This section details some classes of predictive regularity found in Web page classification problems. Are other forms of regularity found in other relational learning domains, or by algorithms different to the one used in my work (FOIL)?