Add comments here for "Categorizing unsupervised relational learning algorithms" by Hannah Blau and Amy McGovern.
From James Cussens: This all made sense to me, and is very much in keeping with the goals of the workshop. It's interesting that you use "link" to denote a relation between objects, as opposed to "relation". The rationale for making this distinction is that in 1st order logic representations (from whence the use of the word "relation" in RDMS) there is no distinction between objects and attribute-values. (Some see this as a problem.) In ILP you might have bond(atom1,atom2) and polarity(atom1,positive) in your background knowledge. Formally, "positive" is supposed to represent some element in some domain, but that's not how we think of it.
From Hannah Blau: Thank you for your comments. We chose the word "link" instead of "relation" because "relation" has so many (perfectly legitimate) senses, it would be easy for a reader to misinterpret what we intended to express. In the area of "relational" learning algorithms, misunderstandings can arise when people use the same word but mean two different things and they don't realize they are not speaking the same language. Misunderstandings like these can occur in any field, but especially in an evolving field where no standard vocabulary has yet emerged.
The ILP knowledge representation is the only one I can think of that treats objects and attribute values uniformly. We hope that the language of object, link, and attribute is useful for talking about other relational learning systems. Of course, there are always alternatives to consider when deciding on a representation for some real-world phenomenon. Sometimes it is not obvious what should be modeled as an object, what as a link, what as an attribute. The same physical reality can yield many different representations depending on which way these decisions go. There is no one correct representation, although some choices are more intuitive than others.