A reason many graph databases don’t offer reasoning is that it is difficult to engineer correctly. Most vendors simply don’t have the skill or expertise required to do it right.
So most of them do an ad hoc hodgepodge of reasoning bits and bobs. At best the attitude is : “Hey, let’s plug in some kind of Prolog engine and hope for the best!” Somewhere after that you get handed a “graph metamodel” and told to do it yourself. Whee! At worst, the vendor spreads FUD about the value of reasoning altogether. Because, yeah, no users could possibly understand Venn diagrams…
Fear, Uncertainty, and Dummies
How do I know it’s FUD? Two reasons.
First, it’s FUD because these same vendors love declarative systems to, say, build software. Their objection isn’t to declarative systems. Their objection is to declarative processing of graphs. Apparently declarative data modeling and analysis systems aren’t fit for their customers.
Second, it’s FUD because in the absence of reasoning, you have to do the work by hand, yourself, anyway and it isn’t easy! Why not let the database figure it out for you instead?
The Whole Truth and Nothing But the Truth…Maintenance
In Stardog reasoning isn’t enabled by default so that our users do not pay the performance penalties if they don’t need the features. Some kinds of “inference” can be performed by using the SPARQL CONSTRUCT form to generate triples that are not explicitly stated in the database. For example, you can add that an individual has an aunt if she has a parent with a sibling who is female or married to a female. Knowing the latter few facts can create the former statement. The constructed triples can be added back to the database and then queried directly.
The problem with this approach is that truth maintenance is a tricky business. If the aforementioned child had an aunt solely because her father’s brother was married, if the uncle gets divorced, she may no longer technically have one. At this point, you would still have the fact that she has an aunt in the database. While it is certainly possible to remove that fact, knowing that you can or should is not usually obvious. If it is stored in the default graph, you may lose track of whether it had been asserted explicitly or derived from a constructed rule.
This is why query-based reasoning is so powerful. Inferred triples can be generated at query time based on the state of the database. Should that change, the inferred results would change, but database administrators don’t have to worry about segmenting or cleaning up inferred statements.
Stardog has always been positioned as the leading graph database for high performance querying and unequaled reasoning capabilities. We are always trying to find ways to improve performance, increase the capabilities, and lower the effort for users to benefit from these features. In Stardog 3.0, we have done all three. We’ll focus on performance improvements in the future, but for now we’d like to draw your attention to two important improvements in Stardog’s reasoning abilities.
Simple Scales > Easy
The first improvement is that users no longer need to know which reasoning level to engage for basic kinds of inference. Most people can benefit from the feature without having to know the difference between QL, RL, EL, RDFS and SL. The distinctions between these levels are important in certain circumstances, but they won’t be for large numbers of uses. To remove this confusion, it is now possible to simply turn reasoning on by setting a flag to “true”. Consult the docs for specifics.
All Things Being Equal…
The other major new addition is full support for OWL 2 equality reasoning. Which means Stardog is the only graph database to offer complete support for OWL 2. Our competitors do have some lovely excuses, however.
So what can you do with equality reasoning? One way you can indicate that two
URIs point to the same resource is to explicitly connect them via the
owl:sameAs property. This suggests that any facts about the one resource are
also true about the other. It sounds simple in practice, but the complexity
owl:sameAs is also symmetric and transitive. This means
that the equality applies in both directions (as it would) and that if A is
owl:sameAs B and B is
owl:sameAs C, then A is
owl:sameAs C. This is a
tremendously useful tool for integrating data sets and it is used widely in the
Linked Data space. In other words, the Berlin mentioned in DBPedia is the same
as the Berlin mentioned in the GeoNames project.
In the end, simpler scales better than easier and reasoning is simpler than doing it yourself, but only when you pick the right database.