Then, the remaining task which is, mind this, conditional pattern base.

For this conditional pattern base, we mine single item patterns,

we construct the FP-tree and mine it recursively.

For example, for p's conditional pattern base, you will get only c: 3,

the reason is f, a, m, and b they never occur

more than three times, so they are out, they are not frequent so you only get 3.

But for m-conditional pattern-base, you per can see f, c, a really

occurs three times, the b only occurred once, then you'll get a f, c, a 3.

And b-conditional pattern-base gives you where you can see none of them,

actually pass the support threshold of 3, so it's empty.

So let's just look at of course, you'll get a and

c's pattern-base like f: 3 and fc:3.

Actually for the single branch, you can dump all the possible combinations there,

support our tree.

But we just look in the recursive way how we can mine this pattern-base.

For this pattern-base, you probably can see, you can say for

a's conditional pattern-base you'll get this.

You'll get fc: 3, and for c you get f: 3, you can see that's the same thing.

Then for this particular am conditional FP-tree you might,

you just take a c, you say c's that means a, m, c or c, a,

m their conditional pattern-base is only f: 3, you get this.

Then you can dump all the patterns like this.

For single branch essentially that's the same thing, you will dump

all the patterns, like all the possible combinations, they are all three, okay.

You can see if you get a single prefix-path,

you actually can partition this into two parts like this path you can mine it,

and this path you can mine it and then you can just concatenate the path and results.

What about if tree cannot fit in the main memory?

If you cannot fit in the memory, we can use database projection.

That means, we project the database based on the patterns,

based on single items set, okay.

So, then we can construct and mine this tree for each projected database.

So we can have parallel projection or

partition projection, two different methods.

The parallel projection means for this one,

you will get f4 predict database and f3 predict database.

For example, for

this first stream you will see gh suppose is non frequent to only f.