[MUSIC] Now, we study another interesting issue called mining negative correlations. So we first need to distinguish rare patterns and negative patterns. What is a rare pattern? Rare pattern usually means there are some rare occurring items, they have very low support but they are interesting. We want to catch such patterns. For example, buying Rolex watches. How to mine such patterns? We previously already discussed this. For different item sets like for those rare items, we should be able to set some individualized group based and minimum supports threshold. That means for rare patterns for just those items, we should set a rather low minimum support threshold, then we'll be able to capture such patterns. But negative patterns could be another very different one. Negative patterns is those patterns that are negatively correlated. That means they are unlikely happen together. So for example, if you find some customer, the same customer, who buys Ford Expedition, which is a SUV car, and also a Ford Fusion, a hybrid car, together. So they are unlikely to happen together, so we called these patterns negative correlated patterns. The problem becomes how to define such patterns? We may have one support-based definition like this. We say, if the itemsets A and B getting together their support is far less than sup(A) x sup(B), that means a chance to get together is far less than random, okay? Then we can say A and B are negatively correlated. Is this a good definition? Actually, this definition may remind us the definition of lift. Then we may see whether they work well for large transaction data sets. Let's look at one example. Suppose a store sold two needle packages, A and B 100 times each, but only one transaction containing both A and B. Then we will see these two needle packages A and B are likely negatively correlated. But when there are in total only 200 transactions in your datasets, you may see s(A U B) getting together, because they got only one time, so 1 over 200, you get this number. This is pretty small number. But then he look at s(A) which is 100 over 200 transaction so it's 0.5. Same as s(B). So their product should be 0.25. So this number is far bigger than this. That means s(A U B) getting together is far less than s(A) x s(B). So we can easily say A and B are negatively correlated, they are negatively correlated patterns. Okay, but when this store, so in total 10 to the power of 5, that means 100,000 transactions. Then suppose all the others does not contain package of A nor B. Then the situation could be completely different, because s(A U B) together is 1 over 10 to the power of 5. But s(A) now is 100 over 100,000, so you get 1 over 1000. s(B) is also 1 over 1000, when they time together you get 1 over 10 to the power of 6. This number is even smaller than A and B getting together. You may say, A and B getting together is very frequent or it's passive correlated, actually it's not. What's the problem? The problem actually is null transactions. Because there are so many transactions that contain neither A nor B, they are null transactions. So we probably can see a good definition of negative correlation should take care of the null invariance problem. That means, when two itemsets A and B are identical related, they should not be influenced. Okay, whether they are an negative correlated or not, they should not be influenced by the number of null transactions. Okay, now we give you another interesting definition, which is a Kulczynzki measure-based definition. That means if we want to say A and B whether they are negative correlated, what we need to see is A and B are frequent. But the condition the probability of A under condition of B and the probability of B under condition of A, their average should be less than epsilon. Where epsilon is a small negative pattern support threshold. Then we probably can see A and B negative correlated can be justified for our needle package problem. We can see no matter there are in total 200 transactions, or 100,000 transactions, if we say epsilon is 0.01, we probably can see this Kulczynski measure based judgement, we can easily see the average of the conditional probability should be less than epsilon, so they are negative correlated. So this seems to be very interesting and a good definition. And how to mine them, actually these are the method similar to our previously discussed pattern mining method. We will not discuss it further. [MUSIC]