Although it's fairly easy to compare two continuous variables and assess their similarity, it's not so straight-forward when you perform the same task on categorical variables. Of course, things are fairly simple when the variables at hand are binary (aka dummy variables), but even in this case, it's not as obvious as you may think.
For example, if two variables are aligned (zeros to zeros and ones to ones), that’s fine. You can use Jaccard similarity to gauge how similar they are. But what happens when the two variables are reversely similar (the zeros of the first variable correspond to the ones of the second, and vice versa)? Then Jaccard similarity finds them dissimilar though there is no doubt that such a pair of variables may be relevant and the first one could be used to predict the second variable. Enter the Symmetric Jaccard Similarity (SJS), a metric that can alleviate this shortcoming of the original Jaccard similarity. Namely, it takes the maximum of the two Jaccard similarities, one with the features as they are originally and one with one of them reversed.
SJS is easy to use and scalable, while its implementation in Julia is quite straight-forward. You just need to be comfortable with contingency tables, something that’s already an easy task in this language, though you can also code it from scratch without too much of a challenge. Anyway, SJS is fairly simple a metric, and something I've been using for years now. However, only recently did I explore its generalization to nominal variables, something that’s not as simple as it may first seem.
Applying the SJS metric to a pair of nominal variables entails maximizing the potential similarity value between them, just like the original SJS does for binary variables. In other words, it shuffles the first variables until the similarity of it with the second variable is maximized, something that’s done in a deterministic and scalable manner. However, it becomes apparent through the algorithm that SJS may fail to reveal the edge that a non-symmetric approach may yield, namely in the case where certain values of the first variable are more similar toward a particular value of the second variable. In a practical sense it means that certain values of the nominal feature at hand are good at predicting a specific class, but not all of the classes.
That's why an exhaustive search of all the binary combinations is generally better, since a given nominal feature may have more to offer in a classification model if it's broken down into several binary ones. That's something we do anyway, but this investigation through the SJS metric illustrates why this strategy is also a good one.
Of course, SJS for nominal features may be useful for assessing if one of them is redundant. Just like we apply some correlation metric for a group of continuous features, we can apply SJS for a group of nominal features, eliminating those that are unnecessary, before we start breaking them down into binary ones, something that can make the dataset explode in size in some cases.
All this is something I’ve been working on the other day, as part of another project. In my latest book “Julia for Machine Learning” (Technics Publications) I talk about such metrics (not SJS in particular) and how you can develop them from scratch in this programming language. Feel free to check it out. Cheers!
Zacharias Voulgaris, PhD
Passionate data scientist with a foxy approach to technology, particularly related to A.I.