WEBVTT 00:01.210 --> 00:02.310 Hi, Dad. 00:02.360 --> 00:08.440 Now, we have learned about how we can play different roles for categorical and numerical variables, 00:08.740 --> 00:12.680 and we have also seen that how decision trees are created. 00:13.270 --> 00:21.970 We know that the decision trees are based on the different rules which we have, and these rules only 00:21.970 --> 00:26.740 define what decision node has to be placed next. 00:27.190 --> 00:34.120 So now we will learn the next most important criteria, which is the splitting of data. 00:35.260 --> 00:46.150 So each split is made would be aligned to an axis with each feature present an axis. 00:46.420 --> 00:47.550 So what does this mean? 00:48.610 --> 00:55.450 So here this means that each split made would be aligned to an axis where each region is in axis itself. 00:55.930 --> 01:02.470 So let us have a look at this particular data where we have values or we can consider these values being 01:02.470 --> 01:03.430 different classes. 01:03.940 --> 01:06.010 So these classes. 01:07.180 --> 01:12.910 I present in four different regions, so the first class is present here. 01:12.960 --> 01:14.120 This is the second class. 01:14.140 --> 01:15.220 This is the third class. 01:15.220 --> 01:16.550 And this is the fourth class. 01:17.140 --> 01:24.340 Now, if they want to subdivide and try to find out where what class is present, then if we want to 01:24.340 --> 01:26.980 make such a decision, then how we can do that. 01:26.980 --> 01:33.510 So we can do that by creating rules so we can create rules by legacy. 01:33.730 --> 01:35.350 Having one rule as. 01:36.460 --> 01:44.650 It's this being the x axis and this being the Y axis, so we can have the first Ruelas X greater than. 01:45.690 --> 01:53.250 Zero point five, when X is greater than zero point five, then this particular line would be selected, 01:53.250 --> 02:00.630 which means that X is greater than zero point five, then it would probably be a part of Class one or 02:00.630 --> 02:01.290 class three. 02:02.470 --> 02:08.830 Now, after we have decided if X is greater than zero point five, then we will have under the condition 02:09.010 --> 02:11.410 that Y is less than. 02:13.060 --> 02:13.970 Zero point five. 02:14.770 --> 02:21.440 So for tree, where we have X greater than zero point five and Y less than zero point five. 02:21.670 --> 02:23.620 That would mean that the glass is one. 02:25.190 --> 02:32.270 Now, remember, the condition we created X is equal to zero point five, X is greater than zero point 02:32.270 --> 02:32.610 five. 02:32.780 --> 02:38.300 So for the condition, yes, the condition which we just discussed about would be prison. 02:39.410 --> 02:48.500 I'm on the other end, we will have X Leiston zero point five, so now for X less than zero point five, 02:48.510 --> 02:50.410 there are two types of conditions present. 02:51.380 --> 03:00.260 So now we will again have another condition that is X is greater than minus one point one. 03:01.650 --> 03:07.030 So if X is greater than minus one point one and this was the previous condition. 03:07.140 --> 03:14.950 So again, the same situation would come up where we will have to create a split on the Y axis. 03:15.930 --> 03:24.180 So based on the y axis, we will again see that if the value of Y is greater than minus zero point five, 03:24.490 --> 03:28.860 if the value is greater than minus zero point five, then it belongs to Ford. 03:29.010 --> 03:33.960 If it is less than zero point five, then there could be a good idea where it could be to. 03:35.180 --> 03:36.400 So we will get to this. 03:37.460 --> 03:45.920 Now, again, there would be another split at this particular point that is being done, Y-axis again. 03:46.160 --> 03:54.140 So it will check if the value is between minus one point two and zero point five. 03:54.440 --> 03:55.940 And it is less than. 03:57.630 --> 03:58.500 Zero point. 03:59.440 --> 04:00.400 Two or three. 04:00.850 --> 04:06.160 And so it will decide that and based on that, it will decide that this would be the last to. 04:07.180 --> 04:16.330 And finally, when the condition was that the value of X is less than or greater than zero point one 04:16.330 --> 04:17.540 minus one point two. 04:17.770 --> 04:23.530 So based on the negatives, that action would be placed for the oil values being for. 04:25.090 --> 04:33.040 So you can see that each split which we are making is in concordance with the X and Y axis, so each 04:33.040 --> 04:37.810 rule which we will be creating here, will be belonging to the X and Y axis. 04:38.730 --> 04:45.040 So let us see how we can actually select different booths so there could be several booths present. 04:45.300 --> 04:52.140 I could have had a room for minus one point five, Wojo I could have had four plus one point five total. 04:52.380 --> 04:57.230 But these rules will actually not give us any good results. 04:57.360 --> 05:01.380 So that is why we have to choose these rules which have been highlighted here. 05:02.290 --> 05:06.980 So how will we actually get to know that Richard Rule is actually useful? 05:07.600 --> 05:12.980 So for that, we will look for a split which gives the most homogeneous chain. 05:13.600 --> 05:18.720 So when we say most homogeneous try node, we don't really know what homogeneous means. 05:19.300 --> 05:21.870 Let us say we don't know the meaning of homogeneous. 05:22.060 --> 05:27.070 So let us try to think about it and what it actually means and let us formulate. 05:27.970 --> 05:35.680 So the different ideas for rule selection are the guinea index, entropy and divisions. 05:35.980 --> 05:40.350 These are three formulas which are used for creating the rules and selecting the rules. 05:41.530 --> 05:47.280 Now, we still don't know how they actually work, but we will look at it in something. 05:48.070 --> 05:56.050 So the Guinea next formula is one minus summation of the probability squit the formula for entropy is 05:56.290 --> 06:01.180 minus probability of probability and the low probability. 06:01.870 --> 06:07.360 So similarly, we can have these formulas in hand and then use them how we will use them. 06:07.370 --> 06:09.120 We will look at it in some way. 06:09.340 --> 06:16.630 So for that, just wait for a moment and let us first of all, understand terrifically what this means. 06:18.870 --> 06:21.480 So let's say we have this entire dataset. 06:22.720 --> 06:31.090 This data contains four glasses, these four glasses of blue, green, red and yellow. 06:32.140 --> 06:38.440 Now, out of these four glasses, I have equal amount of data presenting all the four plus. 06:39.580 --> 06:45.970 Now, the problem here is that I want to decide that what is with. 06:47.460 --> 06:48.900 So let us try to do that. 06:49.590 --> 06:56.980 So for this, I need to create a split in this particular data so that I will be able to recognize what 06:57.090 --> 06:57.900 glasses, glasses it. 06:58.760 --> 07:05.870 If if someone would ask me which a place, a point they were and ask me which point would be that particular 07:05.870 --> 07:09.800 class, I might not be able to answer based on this particular image. 07:11.160 --> 07:14.190 So I need to make a split on this. 07:15.180 --> 07:15.810 Image. 07:17.060 --> 07:19.090 So let us make the split. 07:20.020 --> 07:24.290 So the first bridge that we will be making is the horizontal strip. 07:24.760 --> 07:30.430 So when we make this horizontal split, you can see that data has been divided into two portions. 07:30.430 --> 07:35.130 One is the top portion and the one being the bottom portion in the top portion. 07:35.290 --> 07:45.310 We have us some part of red points, a few blue points and all the yellow points present while in the 07:45.310 --> 07:46.960 bottom split we have. 07:47.990 --> 07:52.430 Some of the great points, all the Green Point and some of the good points. 07:53.540 --> 07:58.120 So here we don't have a clear majority of any points present. 07:59.780 --> 08:05.630 Now, if we have a look at the split, too, you can see they have made this vertical split, which 08:05.750 --> 08:13.670 divides the data into pool spaces where the left hand side space has only to be the point, which is 08:13.670 --> 08:20.540 blue and green, and that to all of the blue and green points right on the right side, we have all 08:20.540 --> 08:22.580 the red and yellow points. 08:23.450 --> 08:31.280 So if you see in this split is done, then if we try to make another split, then what will happen is. 08:32.580 --> 08:33.690 Let us make fun of this. 08:36.320 --> 08:41.600 So for this particular split, the next to which we will have to make is at this point. 08:43.110 --> 08:48.900 Which will divide the blue and green point and the next split which will be created would be at this 08:48.900 --> 08:52.380 point, which will divide the red and the yellow point. 08:52.980 --> 08:58.860 Now, if we talk about the four split and we make another split here, then the split would have to 08:58.860 --> 09:03.730 be made at this particular point and under the split at this point. 09:04.440 --> 09:08.800 Similarly, once it would be made at this place and another the split. 09:09.570 --> 09:16.740 So now if you see they have to make only one, two and three split in the case of the second split, 09:17.430 --> 09:25.260 they've had to make one, two, three, four and five split in the case of the split one. 09:26.070 --> 09:33.050 So it is declining, lesser number of split in case of split two in comparison to split one. 09:33.240 --> 09:40.890 And also the data is being divided in such a way that we have all the data point of one class present 09:40.890 --> 09:41.990 in one position. 09:43.230 --> 09:50.590 So it is better to make a split two in comparison to split one because it divides the data to be in 09:50.610 --> 09:52.640 a more homogeneous location. 09:53.220 --> 10:01.230 So the data in the two is more homogeneous because we get only two classes on the left side and only 10:01.230 --> 10:02.730 two classes on the right side. 10:02.940 --> 10:08.980 While in the case of split one, we have three classes at the top and another three classes at the bottom. 10:09.510 --> 10:13.880 So hence it is less confusing in case of split two. 10:14.100 --> 10:20.520 If I have to make a decision about this at this particular point, then I can easily say on the basis 10:20.520 --> 10:24.240 of this split, which we have made, that I like this place. 10:24.240 --> 10:31.260 Either it would be a blue point or a green point, but if I make the same point in the four split, 10:31.950 --> 10:37.220 then I will have to say that the point could either be blue or yellow orig. 10:37.560 --> 10:41.100 So the confusion which we already had still exists. 10:41.400 --> 10:48.270 We were able to eliminate only one class using this particular split out of the four classes which we 10:48.270 --> 10:53.510 had while we were able to eliminate two classes with just one split in the split. 10:54.640 --> 11:00.250 So that is the reason why we will to now let us discuss what information gain is. 11:01.500 --> 11:08.940 Now for information, Jeanne, let us try to find out the probability first, so for finding out, probably 11:08.940 --> 11:10.980 let us look at this particular data. 11:11.010 --> 11:14.360 So here we have 12 points of data. 11:14.730 --> 11:19.540 So out of these 12 points of data, we are making a split. 11:19.560 --> 11:23.280 I'm driving this down to two different sub nodes. 11:23.730 --> 11:30.750 So the full split is creating this particular division and the second split is making this division. 11:31.720 --> 11:37.650 Now, let us try to find out the probability of the balance for finding the probability of the finding, 11:37.870 --> 11:40.730 we will let us consider the green point. 11:41.110 --> 11:44.620 So what are the probability of getting the green points out of this baneful? 11:44.770 --> 11:48.420 It will be six divided by the total of 12. 11:48.730 --> 11:50.410 That is zero point five. 11:52.000 --> 11:59.590 Now, let us have a look at the probabilities of both the split so far split one at the left hand side, 11:59.770 --> 12:04.330 the probability of getting the green point will be three by the triple. 12:04.330 --> 12:12.010 Seven is equal to zero point for the higher the probability will be three divided by five, which is 12:12.010 --> 12:13.270 zero point six seven. 12:14.660 --> 12:21.600 For the second split, the probability of getting the green pointis four divided by five, which is 12:21.620 --> 12:29.520 zero point B here, the probability is two divided by seven, which is zero point to it. 12:30.350 --> 12:34.240 Now, let us try to find out the probability of getting the Redpoint out of this block. 12:34.580 --> 12:40.690 So that would be five divided by seven, which will be almost 72 percent. 12:41.420 --> 12:48.560 So you can see that when we are trying to decode values from the first split, the four split is less 12:48.560 --> 12:52.310 homogeneous in nature because this is having more mixture. 12:52.970 --> 13:01.040 While in case of the second split, the values are more towards more bias towards a particular class, 13:01.190 --> 13:08.030 and hence we can easily make out that this is a great class study, this is a green class and this one 13:08.030 --> 13:09.250 is a red class. 13:09.890 --> 13:15.680 So when we pick out the point from this particular portion, it will always give out a green point in 13:15.680 --> 13:16.310 majority. 13:16.580 --> 13:19.760 And then we decode points from this particular split. 13:19.850 --> 13:22.510 It will give out more red points. 13:22.790 --> 13:26.750 So hence it being more homogeneous in nature. 13:26.960 --> 13:33.290 We will consider this split as are actual and correct split for the next splitting. 13:35.120 --> 13:38.690 So let us find out the information game using guinea pigs. 13:38.930 --> 13:41.310 So what was the formula for Guinea next? 13:41.320 --> 13:45.520 The formula for guinea pigs was one minus summation of probabilities. 13:45.530 --> 13:51.150 What you can use any of the indexes, you can use Guinea, that you can use entropy, you can use any 13:51.170 --> 13:51.520 means. 13:51.740 --> 13:54.760 But for now, we're explaining it using the Guinea index. 13:56.220 --> 13:59.780 So let us calculate the value for the Fed, including. 14:00.720 --> 14:05.610 This note will have Ghneim Index as one minus. 14:07.080 --> 14:13.590 Six by 12, all squared plus six by 12 square. 14:13.800 --> 14:18.990 So we will be taking the proportion of each class in consideration here. 14:19.200 --> 14:25.950 So it will be one minus six by 12 of the Greenglass and one minus and plus six by 12, all of the grade 14:25.980 --> 14:26.370 class. 14:26.550 --> 14:30.270 And when we solve this, we will get zero point five. 14:31.780 --> 14:35.540 Now, let us get the guinea values for the full split. 14:35.770 --> 14:36.920 This is up for a split. 14:37.120 --> 14:39.580 So the Guinea index would be one minus. 14:40.710 --> 14:41.880 Four by seven. 14:42.890 --> 14:46.340 Plus three by seven all square. 14:46.370 --> 14:49.000 So it comes out to be zero point for evening. 14:49.910 --> 14:56.630 Now let us guys wait for the light inside of the four split, it will be one minus two by five. 14:57.870 --> 14:59.940 Plus three by five. 15:01.090 --> 15:03.760 Holds square and it comes out with zero point forty. 15:04.960 --> 15:08.260 Now we will calculate the weighted average of the game. 15:08.860 --> 15:14.350 Now how will we calculate the weighted average of the weighted average will be calculated by taking 15:14.350 --> 15:15.910 the left hand side. 15:16.090 --> 15:18.110 So how many points we have here? 15:18.130 --> 15:21.510 We have seven points here and in total we have 12 points. 15:21.820 --> 15:29.300 So it will be seven by 12, including gaining of the left hand side, plus the five points here. 15:29.500 --> 15:35.180 So five divided by the of 12 in the beginning of the right hand side. 15:35.320 --> 15:41.970 So it comes out to be zero point forty six and the game comes out to be zero point five dollars million 15:42.010 --> 15:47.860 of the beginning of the BEREC, minus the beginning of the child, which is zero point one for. 15:49.460 --> 15:52.290 Now, let us calculate the need for these second. 15:54.550 --> 16:02.520 The beginning of the second split comes out to be for the left inside the guinea will be one minus four 16:02.710 --> 16:03.220 green. 16:04.350 --> 16:08.850 Divided by the total faith and the whole square of this plus. 16:10.340 --> 16:17.560 One read divided by the total five and one square of this, so it comes out three zero one three two 16:17.570 --> 16:17.920 zero. 16:18.560 --> 16:21.620 Now for the right insight, again, one minus. 16:22.560 --> 16:32.850 Five divided by the total seven plus two, three divided by the total seven and the whole square of 16:32.850 --> 16:36.250 this, so it comes out with zero point four zero eight two. 16:36.900 --> 16:39.090 Now, again, we will take the weighted average. 16:39.360 --> 16:42.430 So this campaign's five out of 12. 16:42.960 --> 16:51.420 So five by 12 into the beginning of the left and say zero point three to zero and plus the. 16:52.600 --> 17:02.040 Seven out of the 12 Endou zero point four zero eight two now we will and we got some of this, will 17:02.040 --> 17:02.170 we. 17:02.500 --> 17:03.940 The two point three seven one. 17:05.500 --> 17:13.000 Now, the gain value will be zero point five minus zero point thirty seven, so it comes out to be zero 17:13.000 --> 17:18.710 point one the week now when we compare the zero point zero one four with zero point one two eight five. 17:19.000 --> 17:25.170 We can easily see that the gain for the second split is more in comparison to the four split. 17:25.360 --> 17:32.020 So we will consider the first the second split for the splitting of this particular pairing. 17:33.130 --> 17:37.540 So we will split a particular node and then we will get these. 17:37.570 --> 17:44.080 So then after getting the split, so once we have split this, then again we will consider how we want 17:44.080 --> 17:49.020 to make the split and which is the best split and again, split this particular node. 17:49.390 --> 17:55.510 So it will keep on going on until we reach a point with each of those conditions, which we have already 17:55.510 --> 17:58.360 set for stopping the splitting. 17:59.880 --> 18:03.420 Now, let us have a look at this particular image. 18:03.630 --> 18:09.550 So here we have the data where we have this these lines of separation. 18:10.110 --> 18:15.230 So let us try to create the rules for this or decision for this. 18:15.630 --> 18:22.480 So the four split, which we can create this X X one less than three point five. 18:23.310 --> 18:27.210 So if X one is less than three point five, then we get this particular region. 18:28.720 --> 18:32.390 And if one is greater than three point five, then we get this particular region. 18:32.410 --> 18:34.720 So let us consider X1 less than three point five. 18:35.710 --> 18:37.970 So the next one is less than three point five. 18:37.990 --> 18:40.010 That is in this particular region. 18:40.300 --> 18:47.620 Then the second condition, which we need to check is if X2 is less than seven point five or not, if 18:47.620 --> 18:53.790 X2 is less than seven point five, then it comes out to be X, otherwise it comes out to be zero. 18:55.520 --> 19:02.240 Now, let us consider the X1 greater than three point five, so if X1 is greater than three point five, 19:02.540 --> 19:06.070 then there could be this particular area which we want to divide. 19:06.910 --> 19:11.960 Now, for this, we need to check if the value is between four four and seven point five. 19:12.110 --> 19:18.410 If the value is between four and seven point five, then we will again check if the value of X1 is less 19:18.410 --> 19:18.980 than seven. 19:20.060 --> 19:25.460 If the value is less than seven and the value is between four and seven point five, then we get this 19:25.460 --> 19:32.140 particular area, which means it is X, and for all other conditions, the market does zero. 19:32.990 --> 19:35.090 So this is how we will create a decision. 19:36.000 --> 19:41.960 Now, once we have got these homogeneous pieces that does this particular space, which is completely 19:41.960 --> 19:49.050 homogeneous and has all the points there is no meaning of creating for the split because the Guinea 19:49.140 --> 19:53.240 index of this particular area will be seen. 19:53.270 --> 19:57.000 The index of this particular area will still be one hundred percent. 19:57.270 --> 20:00.910 And again, if we make another split, it will still be one hundred. 20:00.920 --> 20:03.690 So there will be no information gain present. 20:03.920 --> 20:09.800 So unless and until we have a preliminary information gain, we will not make a further split. 20:10.910 --> 20:18.140 Because if we keep on making up for the split, the three will keep on growing and it will not support 20:18.140 --> 20:24.630 anything, it will just keep increasing the size of the tree without giving much information. 20:27.320 --> 20:32.810 Now, let us talk about regression vs regression trees are what we use in the case of regression. 20:33.020 --> 20:40.810 So in this case, we use the mean and median of the records, I believe not to find out the value, 20:40.820 --> 20:43.040 the average value of the predicted value. 20:43.430 --> 20:47.000 Now, let us try to learn about what regression Tresa. 20:48.130 --> 20:56.440 So for selection of rules in election scenarios, we were using the Guinea Lexar entropy, but in case 20:56.440 --> 21:00.430 of regression, we use the sum of squared error. 21:00.640 --> 21:01.920 So how do we do that? 21:02.110 --> 21:07.180 So we will find out the sum of squared error of the different type of split. 21:08.220 --> 21:13.940 And then find out the minimum split, the value which has the minimum sum of squared error. 21:15.320 --> 21:17.880 So let us consider this particular split. 21:18.530 --> 21:26.330 So this is the original parent or the parent node contains values five, six, four, six, 11, 12, 21:26.330 --> 21:27.320 13 and 12. 21:28.010 --> 21:33.920 Now, if they will protect the values, the predicted value will be the average or the mean of these 21:33.920 --> 21:34.380 values. 21:34.730 --> 21:38.510 So the average or mean of these values comes out to be eight point six. 21:39.200 --> 21:44.690 Now, if we will surprise this eight points, six from the target value, we will get these ordinator 21:44.690 --> 21:45.160 values. 21:45.920 --> 21:50.270 Now we will find out the sum of squared error as described in the. 21:52.550 --> 21:58.970 And the definition of the regression tree, so let us where these values so on squaring these values, 21:58.970 --> 22:02.610 we get thirteen point one seven six point one nine one. 22:02.630 --> 22:05.090 So these are different values of the squared values of the. 22:06.800 --> 22:15.440 And then finally, we are the all of these squared values and get me some of squared error, so what 22:15.440 --> 22:21.890 is almost square and some square in it is we will find out the error, then squared all the error values 22:21.890 --> 22:25.550 and then find us some of these squared error values. 22:27.140 --> 22:32.660 Now we need to make certain splits, and by making the splits, we will have to find out which one has 22:32.660 --> 22:34.900 the lowest sum of squared error. 22:35.890 --> 22:42.700 So this is the first type of split, this four split divides the data such that the target value has 22:42.700 --> 22:49.340 five, six to 11, 12 in the left hand side and 11 six thirty nine four on the right hand side. 22:49.600 --> 22:53.510 So here the predicted value will be the mean of these dot good values. 22:53.650 --> 22:59.320 So for the left side, the mean value is eight point seven five, and for the right side the mean value 22:59.320 --> 23:00.250 is eight point five. 23:01.030 --> 23:06.760 Now we will find out the error by finding the difference between the target and the predicted values. 23:07.180 --> 23:10.720 And then we will square the values and then find the. 23:11.790 --> 23:14.640 Squared error, the sum of squared error. 23:14.850 --> 23:20.350 So now we the add the sum of squared errors of the left and right side of this particular split. 23:20.550 --> 23:25.350 So we get forty two point seven, five plus fifty three, which is ninety five point seven. 23:26.220 --> 23:32.010 And the original sum of squared error of the betting mode was ninety five point eighty seven. 23:32.160 --> 23:36.510 So hence we can see that there is not much improvement by doing this particular. 23:37.050 --> 23:40.500 We have not really gained much from this particular split. 23:41.720 --> 23:47.510 Now there is another split which divides the data such that the target value values on five, six, 23:47.510 --> 23:53.210 six four in the left hand side and 11, 12, 13 and 12 at the right hand side. 23:53.510 --> 23:58.420 Now, again, we will find out the predicted values using the average of these data, good values. 23:58.580 --> 24:01.640 So the average of these would be five point five and four. 24:01.640 --> 24:03.350 These it would be 12. 24:04.390 --> 24:09.460 Now, again, you will find the added value by finding the difference between the target and the predicted 24:09.460 --> 24:09.970 values. 24:10.970 --> 24:17.460 Then we will square these air values and find us some of these squared values. 24:17.750 --> 24:25.130 Now, when we sum the sum of squared errors, we get two point seven five plus two, which is four point 24:25.130 --> 24:25.820 seventy five. 24:26.330 --> 24:29.420 Now, let us compare these sum of squared errors. 24:29.570 --> 24:33.250 So the second squared has a all four point seven five. 24:33.590 --> 24:36.570 The false flag, five ninety five point seven five. 24:36.810 --> 24:42.020 I'm the Fed incommode had a city of ninety five point eighty seven pence. 24:42.140 --> 24:48.770 If we are to choose between different splits, we will decide upon this second split, which reduces 24:49.010 --> 24:53.060 the sum of squared error by almost nine to one. 24:55.240 --> 25:02.530 So this is how we decide on which split we should be choosing now, let us have a look at this particular 25:02.530 --> 25:02.930 image. 25:03.040 --> 25:05.990 This blue line actually shows like what? 25:06.010 --> 25:14.400 How we have created different decision nodes and created split so that we have a decision three in place. 25:15.240 --> 25:22.720 Now, this decision tree is able to predict the training data very greatly and all the values are predicted 25:22.720 --> 25:29.130 very finely, like when we are actually trying to predict certain values, such as the value at this 25:29.440 --> 25:33.790 point and a value at this point or value at this point. 25:33.940 --> 25:38.020 It actually gives a lot of it because for this point it will. 25:38.900 --> 25:43.920 Predict this value for this point, it will predict this value for this point. 25:43.970 --> 25:45.290 This will predict this value. 25:45.590 --> 25:46.400 So now. 25:47.490 --> 25:51.330 If you see the error, which we have is higher. 25:52.270 --> 25:58.990 In these cases, but instead, we could have simply solved this problem by using linear regression. 25:59.560 --> 26:05.380 So if we would have applied linear regression, then we would have created a straight line and it would 26:05.380 --> 26:07.270 have predicted these values could. 26:09.020 --> 26:10.670 Or with the minimal edit. 26:11.580 --> 26:20.460 So this is the reason why we should always start with the very basic models, always implement the linear 26:20.460 --> 26:27.120 regression or logistic regression, forced to get the benchmark value, and then slowly and gradually 26:27.360 --> 26:32.400 get to the next models and see what is the level of improvement we are getting. 26:33.150 --> 26:40.590 If we are not getting much improvement using the higher level of algorithms, then we should always 26:40.590 --> 26:47.220 use the simpler model that is either linear regression or the model which gives a better accuracy. 26:47.970 --> 26:54.960 So that is how we should decide on which model should we be choosing never to go for a complex model, 26:55.170 --> 27:01.780 always decide by having a simpler model which gives a good design to us. 27:02.520 --> 27:09.220 So that is the reason why instead of having a decision, it would have been better to decide and select 27:09.220 --> 27:10.820 the linear regression here. 27:11.580 --> 27:17.550 Like this does not mean that we will always select the linear regression or logistic regression, always 27:17.550 --> 27:20.340 go for a model which gives you a better performance. 27:24.120 --> 27:29.970 Now, we have been creating this decision tree, now this decision tree would have been better if we 27:29.970 --> 27:35.220 would have just dropped earlier point of time, but how would we know that there should we stop? 27:35.700 --> 27:43.440 So to stop, we should either just or select of these five conditions, either of these five conditions 27:43.440 --> 27:45.780 would help us to decide when we should stop. 27:46.140 --> 27:49.430 The first condition being when the node becomes homogeneous. 27:49.710 --> 27:57.300 That is when all the norms of this particular data becomes the same. 27:57.790 --> 28:04.500 When the old data becomes homogeneous, then we should stop splitting because if the nodes are already 28:04.500 --> 28:09.950 homogeneous, then on splitting, there will be no information gain. 28:11.240 --> 28:20.990 Next is when a maximum depth of reporting is reached, so we should decide on that value, a maximum 28:20.990 --> 28:29.630 depth value and based on the maximum value, we can stop at five, seven, 12 or 15 Dep't, which is 28:29.630 --> 28:32.060 considered a good bet for a decision. 28:32.090 --> 28:36.800 Three, and we should not allow the tree to grow a lot. 28:38.170 --> 28:44.950 Next is we should alert, allocate a maximum number of leads, no limit that is reached. 28:45.100 --> 28:52.150 So if we do not know what that we should be having, then we should at least set a limit to the number 28:52.150 --> 28:54.630 of leaf nodes can which can be created. 28:55.380 --> 29:01.720 The next clay is the leaf node has observations less than the loan limit of evolutionism. 29:01.930 --> 29:07.420 So let say I have one hundred rows of data and I am making a particular split. 29:09.460 --> 29:20.050 Now, for this particular split, let's see, I already have the lead up and not having four values 29:20.050 --> 29:24.040 of one class and five values of another class. 29:24.970 --> 29:33.580 Now, if I will be making a split, then it will be creating and giving me one node with forward values 29:33.580 --> 29:36.550 and the node with five values of the other plus. 29:37.520 --> 29:44.450 So this kind of split will be giving me one node which has less than five values. 29:45.420 --> 29:50.250 Which is a very low number of values of No one to muddle through of data. 29:50.970 --> 29:58.740 Now, I could have set the minimum number of Neef, no observations do 20 in case I had a larger amount 29:58.740 --> 29:59.190 of data. 29:59.340 --> 30:03.810 But when I have only hundred rows of data, five would be a lower. 30:04.400 --> 30:08.780 Let's we have a thousand points in the entire data. 30:08.940 --> 30:13.150 Then in that case, we could have said 10 or 20 as the lower number. 30:13.380 --> 30:20.030 So we have to decide on the basis of how much data we actually have so that we don't end up with a leave 30:20.040 --> 30:25.920 node with very minimal amount of data, which does not actually count for a good majority. 30:27.620 --> 30:32.690 Then the split results in at least report having less than no lower limit of value. 30:32.870 --> 30:41.210 So either the D.A. has a lower limit of value or the split after this, particularly if not, will result 30:41.210 --> 30:45.080 in a lower limit of value, then we should stop the splitting. 30:46.750 --> 30:53.470 So there are several hyper barometers, which we have the fullest type of barometer is criterion, which 30:53.680 --> 30:59.500 defines that which method, which way we should be using, which is either guiney or entropy. 30:59.830 --> 31:04.630 The function is to measure the quality of the supported criteria. 31:04.630 --> 31:09.280 Zarghami for the intro before for the information. 31:10.180 --> 31:11.920 Now another is the splitter. 31:12.130 --> 31:16.400 The speaker is either we will choose the best splitter or random speaker. 31:16.600 --> 31:21.040 So which split should you be choosing is what speaker decides. 31:21.060 --> 31:25.260 So by default, the value is best and we should always go for the best. 31:26.110 --> 31:32.830 Next is maximum depth, so maximum depth is the maximum depth of the which we want to have. 31:33.160 --> 31:39.970 So if none of these values, if no value for maximum Deftest chosen, then the Nordon that expanded 31:39.970 --> 31:48.850 till the end of the entire three, till all the leaves are pure or and they all contains less than the 31:48.850 --> 31:49.740 minimum sample. 31:50.800 --> 31:58.030 So we should define a maximum depth or we should define the minimum sample split for. 31:59.670 --> 32:07.620 Next is minimum sample split, minimum sample split has a B fold value to so it will not allow the minimum 32:07.620 --> 32:12.470 sample split to be of value to have more than less than two values. 32:12.690 --> 32:19.050 So the minimum number of samples required to split an internal node is two now. 32:19.050 --> 32:25.060 Next is minimum sample believe that is the minimum number of samples required to be identified. 32:25.360 --> 32:29.120 So there has to be at least one value at the leafleted. 32:29.520 --> 32:32.580 Now, next is maximum features like this. 32:32.730 --> 32:38.910 What is the maximum number of features we should consider by looking for the best split now we could 32:38.910 --> 32:46.050 either choose automatically, choose automatically, or we can choose different values or or square 32:46.050 --> 32:46.840 root or law. 32:46.870 --> 32:53.730 To now maximum feature allows us to select the maximum number of features because we cannot really use 32:53.730 --> 33:01.110 all the features all the time, because then it will create a very huge bucket of rules. 33:01.230 --> 33:04.730 So that is the reason why we don't use a max feature. 33:05.100 --> 33:07.350 Next is maximum leave node. 33:07.530 --> 33:12.060 So it allows to grow a three with maximum leaf node. 33:12.300 --> 33:17.270 That is the maximum number of leaf nodes which could be created. 33:17.880 --> 33:27.780 And lastly, we have last week, which allows us to define if the classes are equally present in the 33:27.780 --> 33:28.770 data or not. 33:29.910 --> 33:31.830 So what does Glass Week help? 33:32.160 --> 33:35.250 So in case, let us say we have a double. 33:35.490 --> 33:37.150 We have two glasses. 33:37.410 --> 33:38.430 Yes, I know. 33:38.820 --> 33:43.380 And there are hundreds of data out of which five hundred rules have. 33:44.280 --> 33:44.820 Yes. 33:44.820 --> 33:48.030 And five hundred rules have no as they're about a good one. 33:48.420 --> 33:53.190 So in this case, we have equal amount of data protection for both the classes. 33:54.180 --> 34:01.590 So hence, we don't need to provide any any details to it, but in case, let us see. 34:01.610 --> 34:09.270 We have a scenario where one glass has only 10 percent of the data and the glass has 90 percent of the 34:09.270 --> 34:18.000 data, then we will have to provide a voyage to the second glass because it has a very less amount of 34:18.000 --> 34:18.390 data. 34:18.780 --> 34:26.550 So what we will do is we will give a high the higher wage to the class, which has less amount of data. 34:26.670 --> 34:33.930 That is the one which has only 10 percent of the data, such that each point of that particular class 34:34.140 --> 34:43.890 is considered as 90 days of the actual or actually nine times of the actual loss so that there are balance 34:43.890 --> 34:45.840 created between the glasses. 34:46.530 --> 34:49.920 So we will be providing that in the way for. 34:52.140 --> 34:56.460 So this is how we will create the maximum split. 34:58.520 --> 35:05.650 Next is Overfitting, next is to define the overfitting, so the decision carries over. 35:05.930 --> 35:12.410 When we have very high depth or there are certain observations which are noisey in nature, and while 35:12.910 --> 35:19.230 we have considered those observations or if we have some features which are not really required and 35:19.250 --> 35:20.620 they are moisy in nature. 35:20.810 --> 35:30.250 So in those situations, overfitting can and thus we have other algorithms which can actually help preventing 35:30.290 --> 35:34.560 overvoting in decision 3s, which we will be launching later. 35:35.540 --> 35:42.710 For now, we will be moving towards the code of decision three and how we can actually implement decision 35:42.710 --> 35:43.110 trees.