1
00:00:01,340 --> 00:00:09,140
As I told you earlier, too, big decision trees with a lot of nodes and a lot of clicks have two problems.

2
00:00:09,730 --> 00:00:16,460
They're difficult to interpret, and they overpowered the training data, thus giving bad asset performance.

3
00:00:18,410 --> 00:00:23,400
Therefore, we decided to control the tree root in the last two religions.

4
00:00:23,720 --> 00:00:30,980
We discussed some methods that stopped tree from growing when a particular condition is made.

5
00:00:33,140 --> 00:00:38,750
These strategies do succeed in giving us smaller and more interpretable trees.

6
00:00:40,480 --> 00:00:41,620
But they have a limitation.

7
00:00:42,850 --> 00:00:45,220
These studies are short sighted.

8
00:00:47,110 --> 00:00:49,780
Suppose we say that we will not make us played.

9
00:00:50,170 --> 00:00:53,890
If it is giving me few units of reduction in MSE.

10
00:00:55,350 --> 00:00:56,950
For simplicity, let's see.

11
00:00:58,320 --> 00:00:59,430
Ten units of benefit.

12
00:01:00,420 --> 00:01:05,070
So if I get 10 units of benefit from the split, I will keep that split.

13
00:01:05,640 --> 00:01:08,010
Otherwise, I will not keep that split.

14
00:01:09,760 --> 00:01:13,730
Now, in this situation, we're at this Naude.

15
00:01:14,290 --> 00:01:21,640
If I make us played, I'm getting a benefit of two units and after this split in this node, if I make

16
00:01:21,640 --> 00:01:24,880
a split, I'm getting a benefit of twelve minutes.

17
00:01:26,710 --> 00:01:27,820
So in this situation.

18
00:01:29,150 --> 00:01:32,120
Because the condition is not met at this node.

19
00:01:33,540 --> 00:01:38,850
Regrowth will be stored started this, not only this node will not be split further.

20
00:01:40,470 --> 00:01:47,130
But if you would have made the split, the next split would have given us the benefit of when we unit.

21
00:01:48,240 --> 00:01:53,140
Two in such a situation will be predesigned the constraint.

22
00:01:54,000 --> 00:02:00,570
We will never find out that this plague was possible because tree growth was stopped at this.

23
00:02:00,570 --> 00:02:01,170
Not only.

24
00:02:02,490 --> 00:02:07,650
Therefore, we need a solution where we do not miss out on such splits.

25
00:02:09,320 --> 00:02:12,150
This solution is called tree pruning.

26
00:02:14,030 --> 00:02:22,640
In this strategy, we draw a very large tree and then prune it, or you can say we cut some parts of

27
00:02:22,640 --> 00:02:27,950
it which are not beneficial in order to obtain a optimal subtree.

28
00:02:29,480 --> 00:02:33,410
So now you order subtree, which is lowest test at a rate.

29
00:02:34,800 --> 00:02:40,290
We can use cross-validation to find out the test at rate of all such Sudbury's.

30
00:02:41,790 --> 00:02:43,650
But this can be computationally expensive.

31
00:02:44,010 --> 00:02:48,240
That is, it may take a very long time to run on this software.

32
00:02:50,280 --> 00:02:56,670
Therefore, we use a method called cost, complexity, pruning or weakest link pruning.

33
00:02:58,180 --> 00:03:05,350
In this matter, we add additional cost of number of terminal nodes in the tree to the Odyssey's.

34
00:03:07,270 --> 00:03:15,010
So instead of minimizing Odyssey's, we minimize odysseys, plus are down for number of terminal node.

35
00:03:16,780 --> 00:03:18,550
Head Alpha is called unique.

36
00:03:18,620 --> 00:03:20,870
But I made it out of complexity better I would it.

37
00:03:23,740 --> 00:03:25,090
If all five zero.

38
00:03:25,840 --> 00:03:27,490
That is normal tree growth.

39
00:03:28,610 --> 00:03:32,540
We are minimizing the normal order system so we get normal re-route.

40
00:03:34,550 --> 00:03:36,470
But the value of a fire increases.

41
00:03:37,730 --> 00:03:41,000
There is increasing penalty for having more split.

42
00:03:45,260 --> 00:03:48,860
Therefore, value of Alpha controls three Grote.

43
00:03:50,560 --> 00:03:58,150
So the process is this we have to find out the value of Alpha at which we get minimum gross validated

44
00:03:58,210 --> 00:03:59,680
error on our training, said.

45
00:04:00,840 --> 00:04:08,700
Using that, many of alpha will prune out tree, and we expect that this prune tree will perform better

46
00:04:08,700 --> 00:04:09,930
on the test set.

47
00:04:12,020 --> 00:04:14,490
So let us see how to do pwning enough software.

48
00:04:14,640 --> 00:04:15,390
Next we do.
