We started this course with the idea that it would be so great if we would be able to learn from examples; get a collection of objects for which we want to predict something (e.g., a collection of people and we want to predict whether someone is healthy or not).
In order to do that we first need to represent these objects. We need to take some measurements, so then we can represent all these objects in terms of these measurements or sensor (feature vector). This would also that we could represent this collection of objects in a scatterplot:
You can see a example scatterplot above, with a dataset of classes A or B.
In previous lectures we studied different classifiers. Now we want to generalize: for new an unknown object (e.g., the green point A in the scatterplot above) what could it be its class label.
Note that other classifiers will also be discussed in the near future:
Recap classifiers
If we have all these classifiers, which one should one use? You could try them all, and then pick the one with the lowest error (applying the result of the training set to the test set). However, this is brute and inefficient, specially if we are limited in time and power.
Sadly this is something we need to in most cases… However, there are some characteristics of the data sets which may give us some clue of what classifier to use.
Take for example:
As the number of data points is not normally large, using Gaussian-based classifiers may give a lot of issues (as discussed previously). The same intuition can be used with the others classifiers, so we do not need to try all of them.
Let’s see a typical situation:
In real life, data is not 2D beautiful, but we have high-dimensional data (in this case, 150 objects in 4D), making it even harder to choose a classifier. Unfortunately, in this case we can only brute-force and try all of them.
But, how do we choose what classifier is better? We need some objective criterion. Typically: classification performance/error.
Note that: