WEBVTT 00:01.030 --> 00:09.220 In this session, we will discuss about the forced ensemble learning method, which is bigging, which 00:09.220 --> 00:13.090 has been implemented in the form of random forest. 00:14.920 --> 00:16.090 So let us look at it. 00:17.180 --> 00:25.880 So first of all, let us discuss about the session, please, so large decision trees have low bias 00:26.000 --> 00:30.020 and high variance and they tend to overthink. 00:31.300 --> 00:39.550 While in case of storms, the storms have high bias and low ratings and depend on the. 00:40.950 --> 00:54.150 Now, these will be creating all random forest, random forest is an implementation of bagging and bagging 00:54.300 --> 00:57.700 helps us to reduce the variance. 00:57.960 --> 01:06.690 So how do we do that is we basically combine different stumps and we take the average of them so that 01:06.690 --> 01:09.650 we get the low variance model. 01:10.170 --> 01:14.190 So let's see for the what are we doing in case of body? 01:14.490 --> 01:17.850 In case of bug, we reduce the variance. 01:18.390 --> 01:25.330 So we combine the stumps so that the audience is reduced in comparison to the degree. 01:25.350 --> 01:29.510 So in case of decision tree, we have high variance. 01:30.060 --> 01:37.160 So what we do is we basically use stumps so that we could have low variance in this particular situation. 01:39.220 --> 01:42.740 So how what is actually random forest? 01:43.510 --> 01:50.320 So we have discussed about bigging, we have discussed about why they use the stumps instead of decision 01:50.320 --> 01:53.860 trees, but we don't really know what random forest is. 01:55.350 --> 02:06.090 So random forest is a group of small decision trees, that is a group of stumps which we think and using 02:06.090 --> 02:12.480 these group of stumps, we create a huge forest, we create a forest. 02:12.480 --> 02:13.410 What is a forest? 02:13.410 --> 02:16.320 Forest is something that contains a lot of trees. 02:16.710 --> 02:22.640 So same thing is here we are creating forest of decision stumps. 02:23.160 --> 02:29.720 Now, when we are creating the forest of decision brown stumps, we need to add this randomness. 02:30.300 --> 02:38.220 So when we are seeing random forest, it means that we are creating a forest and we have to add some 02:38.220 --> 02:39.490 randomness to it. 02:40.810 --> 02:50.440 Now, how are we introducing this randomness, this randomness has been introduced by creating subsect 02:50.440 --> 02:54.940 of samples, which is taken for the framing of each and every three. 02:56.080 --> 03:03.160 So when you will be creating that decision, what we do is we invite the decision three, we consider 03:03.160 --> 03:10.330 all the features which are present now in case of random forest, when we will be creating the decision 03:10.330 --> 03:18.490 stems, we will be taking the small samples of data instead of considering all the samples of data. 03:19.000 --> 03:27.240 So we will take only small samples of data so that we can get only a subset of the data points. 03:27.400 --> 03:30.970 That is only a subset of patterns from the data. 03:32.440 --> 03:40.240 The next thing what we do is we also take a subset of the variables by draining each tree. 03:40.870 --> 03:44.620 Now, what happens when we are taking the subset of variable now? 03:44.620 --> 03:47.340 Because we are taking a subset of variables. 03:47.500 --> 03:56.440 So instead of having all the features present for creating a tree, we are now having different subsets 03:56.440 --> 03:58.960 of features for creating that decision. 03:59.860 --> 04:07.950 Now, imagine if we would have used the complete set of features for creating the decision tree stumps. 04:08.170 --> 04:13.690 Then what would have happened is that because it will be evaluating all the root? 04:14.910 --> 04:22.080 And because it is evaluating all the rules on the basis of same Ghneim, they sort of squared errors. 04:22.300 --> 04:32.350 That is the reason why it will be the next thing, the rules which are having the highest Guiney gain. 04:33.090 --> 04:40.830 So that is the reason why it was the same split in the first and then followed by the next split. 04:41.040 --> 04:47.100 So the decision tree, which will be formulated will be just the same because there is no new feature 04:47.100 --> 04:54.090 or no subset of feature, which is why when the new features will not be there, when the old features 04:54.300 --> 05:01.440 will not be accepted, then it will be creating exactly the same decision because there is no variation 05:01.440 --> 05:01.760 in it. 05:02.370 --> 05:07.560 So that is why we need to provide the variation in the form of randomness. 05:08.630 --> 05:15.650 Which is the reason why we are looking at a sample subsid so from the entire data, we will look at 05:15.650 --> 05:18.400 small samples of the data at every time. 05:19.010 --> 05:26.020 And for creating the rules, we will be creating all the rules from all the variables. 05:26.030 --> 05:29.390 We will not be considering all the variables for creating the three. 05:29.690 --> 05:34.140 We will be creating only rules from the small subset of the variables. 05:34.640 --> 05:40.910 Now, again, for each and every split which we will be creating, the rules are not considered from 05:40.910 --> 05:42.540 the all in variables. 05:43.070 --> 05:44.090 So let us see. 05:44.270 --> 05:50.630 Initially, we have a number of variables from which we use to create our decision. 05:50.630 --> 05:52.580 Three these rules. 05:52.640 --> 05:54.680 But all the variables which we had. 05:55.840 --> 06:05.410 So now for creating the stump one stump to stump three and four, I will be taking small subsets of 06:05.410 --> 06:06.900 these variables. 06:07.240 --> 06:11.440 So now the subsequent have any number of variables. 06:11.740 --> 06:16.740 So each more they created from any number of variables. 06:16.960 --> 06:25.300 I like creating the subsect while creating this small stump out of these and variables out of these 06:25.540 --> 06:26.290 variables. 06:26.560 --> 06:30.640 Each split, which I will be making now, the rules will be big. 06:30.640 --> 06:35.890 No, from these in variables, but again, a subset of these in variables. 06:36.880 --> 06:37.480 So you see. 06:38.490 --> 06:46.590 The four subsegments we are taking is like creating the tree itself by creating the stump itself or 06:46.590 --> 06:55.530 the Vitullo itself, so the vehicle itself could be created from a subset of the original number of 06:56.010 --> 07:02.760 features which we had from these variables, which we now have the smaller subset of people, which 07:02.760 --> 07:10.330 we now have from this also every split which we will be making at every split. 07:10.350 --> 07:15.240 Again, we will not consider all the rules, but only a subset of these rules. 07:16.960 --> 07:26.710 So this will again be the one the miss in the stump or in the vehicle, which we will be using now as 07:26.710 --> 07:32.580 a random subset of features selected for the --, are different for each and every split. 07:32.920 --> 07:34.980 So this is what we have known for now. 07:35.110 --> 07:38.620 So that is that the subset of the samples will be done. 07:38.800 --> 07:45.580 Then the subset of the variables will be considered for creating the pre and after creating these three 07:45.580 --> 07:46.780 from this subset. 07:47.800 --> 07:56.230 And every split, again, we will be considering a subset of the number of variables which we have now. 07:58.360 --> 08:03.030 So this is how we will be adding randomness to the decision to. 08:09.420 --> 08:13.350 Now, what a different type of enemy does we have for them? 08:13.380 --> 08:15.990 What is the hypovolemic those? 08:17.330 --> 08:26.060 The investigators, what are the investigators these are in these numbers, which is the which is the 08:26.060 --> 08:34.480 number of bodies which we will be creating so we can create the one hundred three point five hundred 08:34.490 --> 08:37.730 these thousand trees, any number of trees. 08:38.060 --> 08:43.190 But the prescribed number is five hundred five hundred usually works. 08:43.190 --> 08:50.090 Well, in the case of random forest, we usually try a hundred five hundred thousand and then narrow 08:50.090 --> 08:51.110 it down for the. 08:52.880 --> 09:01.880 Next is maximum depth, maximum bet is again in these informal eye, this shows the maximum depth of 09:01.880 --> 09:06.110 the model or the vehicle on which we will be using. 09:06.290 --> 09:12.950 So each week, lonas, that is the maximum that usually we keep it around to five. 09:13.820 --> 09:15.500 Then we have Mike Sumfest. 09:16.860 --> 09:23.940 This is the number of samples to draw from the X Supreme E to base estimated that this. 09:25.730 --> 09:31.820 This maximum number of samples, so what is the maximum number of samples which we have to select? 09:32.670 --> 09:42.330 For creating the estimated then minimum impurity degrees, that is an order will be split if this split 09:42.330 --> 09:49.230 induces or degrees of the impurity greater than the equal of this value that this. 09:51.230 --> 10:00.260 That is the minimum amount of impurity or degrees, or you can also see information gain the information 10:00.260 --> 10:07.460 gained, which we were talking about, we said that if we have a fixed amount of information gain at 10:07.460 --> 10:12.460 least a minimum amount of information gained, then only we will make a split. 10:12.710 --> 10:14.900 Otherwise we will not make a split. 10:15.920 --> 10:22.970 Then warm, stop it, if it is said to be true, then it will allow us to refuse the solution of the 10:22.970 --> 10:30.830 previous goal of the Fichte function and it will allow us to add more estimated dollars to the ensemble. 10:30.830 --> 10:34.400 Otherwise it will just fit a whole new forest. 10:36.410 --> 10:42.830 Then in the jobs is what defines the number of jobs which begin the running buddy. 10:43.070 --> 10:50.210 So instead of creating the number of threes sequentially, we can actually create them. 10:50.210 --> 10:53.090 Barlinnie by using any jobs. 10:56.680 --> 11:00.680 So this is the theory about random forest. 11:01.120 --> 11:04.900 Next, we will learn about the implementation of random forest. 11:08.450 --> 11:15.360 Now, at this point of time, you have a good knowledge of how a model should be generated. 11:15.800 --> 11:20.420 So now you can actually start to create your project. 11:21.710 --> 11:30.710 So to create your project, you can become the B2C provider and to the data set, you can now actually 11:30.710 --> 11:34.270 run the NINIAN models and often linear models. 11:34.400 --> 11:38.450 You can start implementing decision tree, random forest. 11:38.570 --> 11:47.720 And as we go ahead, you will be implementing stocking and extra boost to it and finally create a pipeline 11:47.870 --> 11:50.760 of the best models which you have generated. 11:51.710 --> 11:58.670 This is a project which will allow you to try all the aspect of machine learning. 11:58.880 --> 12:07.280 That is from data preparation to data analysis and also the implementation of various machine learning 12:07.280 --> 12:08.030 algorithms. 12:08.210 --> 12:09.590 And it will allow you. 12:10.590 --> 12:17.990 To compare different machine learning algorithms with each other so that you can easily create the models 12:18.000 --> 12:23.220 when you want to create them in the real life during your walk.