So the examples we've talked about so far in this class, you often know what the labels are. In other words, you're trying to do supervised classification, you're trying to predict an outcome that you know what it is. But sometimes you don't know the labels for prediction, so you have to discover those labels in advance. So, one way to do that is to create clusters from the data that you've observed, to add names to those clusters and then build a predictor for those clusters. In a new data set, you're going to predict the cluster and apply the name that you come up with in the previous data set. This adds several levels of difficulty to the prediction problem. First of all, creating the clusters is not a perfectly noiseless process. Second of all, coming up with the right names, in other words interpreting the clusters well is an incredibly challenging problem. And if you've taken exploratory data analysis in our data science specialization, you'll understand why. Building a predictor for the clusters is basically using the algorithms that we've learned throughout the course of this class, as well as predicting on those clusters. So I'll just show you a quick example of how you could do this. So, if I load the iris data and the ggplot library, I can then create a training and test set for the iris data just like I have in previous examples, but I could ignore the species clusters. So one thing that I could do is I could perform actually a k-means clustering. If you remember that k-means clustering from the exploratory data analysis section of the data science specialization. If you don't, you can look it up on Wikipedia. And the basic idea here is to basically create three different clusters. So I was telling it to create three different clusters, ignoring the species information. So then what I can do is, I can basically assign to those clusters different colors. And so here I've plotted those three different clusters that are created by performing this k-means clustering, so there's a cluster here, a cluster here and a cluster here. Those clusters actually are relatively close to the clusters you would get if you used the species labels themselves, but that's not a typical outcome, often it's very hard to label the clusters that you get. So in this case, I can make a table of the cluster versus species, and I can see that cluster 2 corresponds to this species. Cluster 3 corresponds to this species and cluster 1 corresponds to this species. So in general, I wouldn't know what those species names were, and I would have to come up with names for each of my clusters. And then I can fit a model that relates the cluster variable that I've just created, to all the predictor variables and I could do it in the training set. In this case, I am doing it with a classification tree. I can then do a prediction in a, a training set, and I can see that I do a reasonably good job of predicting this cluster and this cluster but cluster 1 and cluster 3 sometimes get mix, mixed up in my prediction model. That's because I have both error or variation in my prediction building and error and variation in my cluster building, so it ends up being a quite a challenging problem to do, unsupervised prediction in this way. I can then apply it on the test data set. So if I predict on the test data set, in general I wouldn't know what the labels are, but here I'm showing what the labels are. So, here I'm predicting on a new data set and making a table versus the actual known species. And so I can see this actually does quite a reasonable job here of predicting the different species into different clustered labels. In general, this is quite a hard problem, so care must be taken in the labeling the clusters, and performing the unsupervised analysis and understanding how that works. The cl_predict function in the clue package provides similar functionality to what I've just described here, but in general it makes a little bit more sense to sort of build your own approach if you're doing unsupervised prediction because you really need to think carefully about how you're going to define the clusters. Be very wary of over interpretation of this type of clusters. This is in fact an exploratory technique. And so the clusters may change depend on the way that you sample the data. This is actually one basic approach to building things like recommendation engines, where you identify clusters of people that have similar tastes and assign their tastes to new individuals. You can read more about unsupervised prediction in the elements of statistical learning and introduction to statistical learning.
