Alright, so last time I went over what was called Structural Hamming Distance. It’s basically a measure for comparing how good a graph is compared to another graph. However, it has some flaws, mainly ones rooted in its simplicity. The goal behind a scoring metric is that it’s able to tell us the most amount of information in as little outputs as possible. SHD outputs one number, but that’s basically just how many flaws it finds in a graph. It doesn’t say anything about the severity of the mistakes or even what those mistakes are. Enter Structural Interventional Distance, or SID for short. This metric was developed with the idea that any network graph can be broken down into smaller sub-graphs (markov classes, anyone?). So, taking into account these smaller graphs, we can visualize the relationships using do-calculus, and then report how many of these interventional probabilities match up. Here’s an example:
Again, this doesn’t necessarily tell us exactly which parts of the graph are wrong on its own; however, it does create a more scalable system for just HOW wrong a graph’s anomalies are. Which is cool, and useful. Especially when used in conjunction with SHD – if two graphs had the same SHD, we know how many mistakes there are, but if one SID is significantly higher than another, then we know that the mistakes in that graph are much more severe. Here’s to hoping I can do something with that in my test programs.
Searching For a Cause 