1.14 Machine learning lecture 15  (Page 4/15)

Let's just delve a little deeper into those examples to convey more intuition about what LSI is doing. So look further in the definition of the co-sine similarity measure. So the numerator or the similarity between the two documents was this inner product, which is therefore sum over K, XIK, XJK. So this inner product would be equal to zero if the two documents have no words in common. So this is really – sum over K – indicator of whether documents, I and J, both contain the word, K, because I guess XIK indicates whether document I contains the word K, and XJK indicates whether document J contains the word, K.

So the product would be one only if the word K appears in both documents. Therefore, the similarity between these two documents would be zero if the two documents have no words in common. For example, suppose your document, XI, has the word study and the word XJ, has the word learn. Then these two documents may be considered entirely dissimilar.

[Inaudible] effective study strategies. Sometimes you read a news article about that. So you ask, what other documents are similar to this? If there are a bunch of other documents about good methods to learn, than there are words in common. So similarity [inaudible]is zero.

So here's a cartoon of what we hope [inaudible] PCA will do, which is suppose that on the horizontal axis, I plot the word learn, and on the vertical access, I plot the word study. So the values take on either the value of zero or one. So if a document contains the words learn but not study, then it'll plot that document there. If a document contains neither the word study nor learn, then it'll plot that at zero, zero.

So here's a cartoon behind what PCA is doing, which is we identify lower dimensional subspace. That would be sum – eigen vector, we get out of PCAs. Now, supposed we have a document about learning. We have a document about studying. The document about learning points to the right. Document about studying points up. So the inner product, or the co-sine angle between these two documents would be – excuse me. The inner product between these two documents will be zero. So these two documents are entirely unrelated, which is not what we want.

Documents about study, documents about learning, they are related. But we take these two documents, and we project them onto this subspace. Then these two documents now become much closer together, and the algorithm will recognize that when you say the inner product between these two documents, you actually end up with a positive number. So LSI enables our algorithm to recognize that these two documents have some positive similarity between them.

So that's just intuition about what PCA may be doing to text data. The same thing goes to other examples and the words study and learn. So you have – you find a document about politicians and a document with the names of prominent politicians. That will also bring the documents closer together, or just any related topics, they end up [inaudible] points closer together and just lower dimensional space.