1.5 Machine learning lecture 6  (Page 17/15)

MachineLearning-Lecture06

Instructor (Andrew Ng) :Okay, good morning. Welcome back. Just one quick announcement for today, which is that this next discussion section as far as for the TA’s will mostly be on, sort of, a tutorial on Matlab and Octaves. So I know many of you already have program Matlab or Octave before, but in case not, and you want to, sort of, see along the tutorial on how direct terms and Matlab, please come to this next discussion section.

What I want to do today is continue our discussion of Naïve Bayes, which is the learning algorithm that I started to discuss in the previous lecture and talk about a couple of different event models in Naïve Bayes, and then I’ll take a brief digression to talk about neural networks, which is something that I actually won’t spend a lot of time on, and then I want to start to talk about support vector machines, and support vector machines is the learning algorithms, the supervised learning algorithm that many people consider the most effective, off-the-shelf supervised learning algorithm. That point of view is debatable, but there are many people that hold that point of view, and we’ll start discussing that today, and this will actually take us a few lectures to complete.

So let’s talk about Naïve Bayes. To recap from the previous lecture, I started off describing spam classification as the most [inaudible] example for Naïve Bayes in which we would create feature vectors like these, right, that correspond to words in a dictionary. And so, you know, based on what words appear in a piece of email were represented as a feature vector with ones and zeros in the corresponding places, and Naïve Bayes was a generative learning algorithm, and by that I mean it’s an algorithm in which we model PFX given Y, and for Naïve Bayes, specifically, we modeled it as product from I equals one to N, PFXI given Y, and also we model PFY, and then we use Bayes Rule, right, to combine these two together, and so our predictions, when you give it a new piece of email you want to tell if it’s spam or not spam, you predict RFX over Y, PFY given X, which by Bayes Rule is RFX over Y, PFX given Y, times BY, okay?

So this is Naïve Bayes, and just to draw attention to two things, one is that in this model, each of our features were zero, one, so indicating whether different words appear, and the length or the feature vector was, sort of, the length N of the feature vector was the number of words in the dictionary. So it might be on this version on the order of 50,000 words, say.

What I want to do now is describe two variations on this algorithm. The first one is the simpler one, which it’s just a generalization to if XI takes on more values. So, you know, one thing that’s commonly done is to apply Naïve Bayes to problems where some of these features, XI, takes on K values rather than just two values, and in that case, you actually build, sort of, a very similar model where PFX given Y is really the same thing, right, where now these are going to be multinomial probabilities rather than Bernoulli’s because the XI’s can, maybe, take on up to K values.