WEBVTT 00:01.140 --> 00:08.860 Now, let us have a look at this and type of item does, which we have in case of SUV, in case of SVM. 00:09.150 --> 00:13.460 We will be working with several hypovolemic that the government. 00:14.280 --> 00:18.810 So what is going on is a mathematical function. 00:19.710 --> 00:27.150 This function of government is to the data as input and transform it into a required form. 00:27.510 --> 00:27.990 So. 00:29.590 --> 00:37.240 Here you can see they have this data, which is in the form of the data, which is in two dimensions. 00:37.600 --> 00:45.070 And what now, what Cornell would do is government would transform this data from these two dimensions 00:45.280 --> 00:52.340 to three dimensions where it will add another dimension to it and move earlier. 00:52.510 --> 00:59.980 The problem was that because X and Y were two different dimensions and the data had no linear separation. 01:00.170 --> 01:05.290 So we were not able to separate these and we were not able to create a line between this. 01:06.380 --> 01:13.550 Now, with the presence of the colonel, what happens is the coroner transforms this data from this 01:13.550 --> 01:16.370 dual dimensions to three dimensions. 01:17.060 --> 01:27.500 Now, these three dimensions actually allow us to create a plane between X in Z axis, and this plane 01:27.650 --> 01:31.580 is able to separate both of these clusters. 01:32.420 --> 01:40.460 Now, when we transform this data back to the linear plane, back to the X and Y plane, we can see 01:40.610 --> 01:47.270 that the line, the plane, which we had all actually created, had formed the circle, the plane. 01:48.230 --> 01:54.530 Which is able to classify X and Y and separate X and Y clearly. 01:57.340 --> 02:04.630 Now, what is the coroner the coroner defines the distance measure between new data and the support 02:04.630 --> 02:12.310 vectors that this observation poses to the hyperbole in case of a corner, the corner will be defining 02:12.310 --> 02:14.880 the distance between the new data. 02:15.460 --> 02:22.780 So the government will make sure that the new data, which we are adding to this particular data, the 02:22.780 --> 02:30.880 data which we are actually testing is finite, is evaluated with respect to the government, with respect 02:30.880 --> 02:32.480 to the distance from the plane. 02:32.740 --> 02:39.730 So it is able to define the distance and based on that, it can actually find out which observation, 02:39.730 --> 02:43.060 which are closest to the type of plane and which are not. 02:44.250 --> 02:51.510 The values which are closest to the high poverty, they are known as the support workers and based on 02:51.510 --> 02:56.010 the support vectors, it actually defines how the classification needs to be. 02:57.240 --> 03:05.410 Now, the main task here is how far or how close the lines could be to these support vectors. 03:06.270 --> 03:09.660 That is the hyperbole in which we are creating. 03:09.840 --> 03:14.550 How far or how close can we allow them to get through this support mechanism? 03:15.030 --> 03:23.130 The closer the separation line will be to this support vector, the more accurate the. 03:24.780 --> 03:27.120 Classification will be back. 03:27.790 --> 03:33.740 Another thing is that there are a few more parameters which allow us to define these things. 03:34.140 --> 03:36.230 So let us have a look at that now. 03:36.240 --> 03:46.680 And during the hyper the higher dimensions, the corners will allow the data to be classified in a higher 03:46.680 --> 03:47.290 dimension. 03:47.460 --> 03:52.980 So there are corners which are present for the lower dimensions of their corners, which are present 03:52.980 --> 03:54.320 for the higher dimensions. 03:54.570 --> 03:59.750 So the corner like polynomial corner and a radial corner. 03:59.910 --> 04:04.350 These are two corners which are used for the higher dimension. 04:04.530 --> 04:08.220 So they transform the input space into higher dimension. 04:08.400 --> 04:13.400 If they have a linear kernel, it will not transform the data into higher dimension. 04:13.410 --> 04:16.530 It will just walk on the X and Y axis. 04:16.740 --> 04:20.790 But when we have the polynomial going, it will only. 04:21.960 --> 04:27.850 Wolf, it will allow to have more complex functions and considered more complex functions. 04:28.100 --> 04:35.250 Eileen O'Connor will only create a straight line so a linear kernel will only be able to create a straight 04:35.250 --> 04:45.000 line via the polynomial and radial kernel will be able to capture the complex features, also the complex 04:45.000 --> 04:46.440 shapes of the data. 04:46.920 --> 04:51.590 So that is why for higher dimensional data, we use polynomial and. 04:54.830 --> 05:02.030 Now, here we can see the visualization of the scene, so here when we have a linear kernel, we have 05:02.300 --> 05:05.550 a straight line of separation which is needed. 05:05.960 --> 05:13.540 So here we have just created a straight line and the straight line was able to classify the data accurately. 05:14.330 --> 05:21.540 But when the data is actually in a polynomial form, that is, it has different goals. 05:21.740 --> 05:31.250 So in that case, we can use the polynomial and the situations where the 5N is a lot more complex and 05:31.250 --> 05:33.500 cannot be figured in a single goal. 05:33.860 --> 05:38.740 So in those cases, we will be using these radial basis function. 05:39.620 --> 05:46.700 Here you can see that here we have small clusters of data separator and these are still belonging to 05:46.700 --> 05:47.660 the same classes. 05:47.660 --> 05:50.240 And around these we have the. 05:51.360 --> 05:57.270 Another class, so these are more accurately classified by the radial basis function. 05:58.540 --> 06:07.000 Now, how far these line of separations will be is decided by the hyper barometer. 06:07.990 --> 06:15.650 So that is see value now what they see value, see value is on the regularization parameter. 06:15.940 --> 06:18.670 Now, when the value of C is large. 06:20.080 --> 06:29.170 Then the value of sea is large, the smaller margin hyper plane will be considered since it stresses 06:29.320 --> 06:38.470 on getting all the brownie points classified vertically and a small value of sea will consider a large 06:38.470 --> 06:40.180 margin piper plane. 06:40.180 --> 06:44.630 And even if some points are misclassified by the hyperbole. 06:44.920 --> 06:49.600 So what does the sea value emphasize on the sea value? 06:49.600 --> 06:54.340 Emphasis on the distance from the support vectors. 06:54.790 --> 06:55.240 So. 06:57.360 --> 07:02.140 Here we have these vectors and here we have this hyperbole. 07:02.760 --> 07:11.010 Now, if the value of SI will be large, then we will be trying to have the distance smaller and then 07:11.010 --> 07:13.750 the value of SI will be small. 07:13.770 --> 07:19.400 Then we will try to maximize the distance from this site separating hyperbole. 07:19.770 --> 07:22.410 So let's have a look at the image related to that. 07:22.590 --> 07:27.510 So here you can see when sea value is very small, the distances get very high. 07:29.120 --> 07:32.630 So when the distances get very high, the. 07:33.700 --> 07:40.630 Few of the data point actually fall inside this and these might get misclassified. 07:42.320 --> 07:51.080 Why then did we increase the value of see, what happens is that the margin is kept very small and so 07:51.080 --> 07:55.230 that there is no point which gets misclassified. 07:55.400 --> 08:00.780 So the larger the value of C, the lesser the chance of misclassification. 08:01.640 --> 08:04.400 So here you can see the diagram later. 08:04.440 --> 08:06.730 So here we have created a clear line. 08:07.130 --> 08:07.910 I'm here. 08:08.090 --> 08:15.560 The value as the value of these is that more closely to the data classes, which we have. 08:16.880 --> 08:24.530 So the higher the value of C, the lesser the chances of misclassification, because then the margin 08:24.530 --> 08:26.570 separation will be less than. 08:28.680 --> 08:32.470 And a little bit by the media is the guerma value. 08:32.670 --> 08:40.890 Now, the gamma barometer defines how far the influence of each training observation affects the calculation 08:40.890 --> 08:42.710 of the optimal hyperbole. 08:43.080 --> 08:50.700 That is how far the influence of each training observation affects the calculation of optimal hyperbole. 08:51.000 --> 08:56.380 Now it defines how far the influence of a single training example reaches. 08:56.550 --> 09:04.330 So if they have one training example, then how much influence will it have on the hyper creation? 09:04.560 --> 09:11.370 If that is, each point will be having the area around it, which will be impacted. 09:12.590 --> 09:21.920 So if they have low value of Gamma, then it will then up Datapoint will have a very long impact for 09:21.920 --> 09:23.620 the hypovolemic creation. 09:23.870 --> 09:26.570 That is, it will have a father boundry. 09:26.750 --> 09:34.730 When we have a high value of gamma, it means that we want to have the type of plain-clothes it will 09:34.730 --> 09:37.650 allow the hyper plane to be created closer to it. 09:38.450 --> 09:43.090 Now the gamma barometer is the inverse of the radius of the influence. 09:43.280 --> 09:52.490 So the higher the value of gamma, the lower the radius will be lower the radius of the lensmen. 09:53.150 --> 09:53.600 So. 09:55.820 --> 10:03.890 Ladies, so here we have Gameau values small, so this means that the radius of influence will be high, 10:03.890 --> 10:07.640 so it will try to keep the boundary as far as possible. 10:07.880 --> 10:13.610 That is why the boundaries are straight in line and are farther away from the point. 10:14.000 --> 10:22.270 Why, as the value of gamma increases, the points allow the boundary to be closer to them. 10:22.550 --> 10:30.770 So because the impact of the point is low, so it is allowing the type of hydroplaned to be closer to 10:30.770 --> 10:30.910 it. 10:31.160 --> 10:33.830 Hence we have such boundaries. 10:36.290 --> 10:45.850 Now, let us have a look at the pros of this swim so it works really well with a clear margin of separation. 10:46.190 --> 10:55.190 So in case we have data which has a clear space of margin where the classes are well separated in those 10:55.190 --> 10:55.910 conditions. 10:57.060 --> 11:06.060 SVM will work very well next is it is effective in higher dimensional spaces because we have the corners 11:06.060 --> 11:13.740 available which can actually deal with high dimensions, then it is effective in cases where the number 11:13.740 --> 11:17.370 of dimension is greater than the number of samples. 11:18.240 --> 11:19.320 Why is this so? 11:19.500 --> 11:27.150 Let's say we have 20, 20 data points and we have more number of dimensions. 11:27.490 --> 11:31.590 Number of dimensions would be the number of features that we have now. 11:31.710 --> 11:38.700 If we have more number of features, there will be more number of dimensions available now because while 11:38.700 --> 11:46.620 classifying using SVM, we are not really worried about all the data points which are present, but 11:46.620 --> 11:50.340 only about four years hence. 11:50.550 --> 11:53.570 We don't really need that much amount of data. 11:55.070 --> 12:01.760 And SVM, there is no need to have large amount of data, would have a better performance for having 12:01.760 --> 12:07.370 a better performance, we need a better, cleaner separation than what we are going to have a better 12:07.390 --> 12:14.000 performance, but more a little bit more number of rules of data that would not really mean that we 12:14.000 --> 12:17.380 can have a better classified from the film. 12:17.630 --> 12:25.970 OK, so in this case, if it can work really well, if that is high amount of dimensions, because if 12:25.970 --> 12:31.940 there are a number of dimensions, then it will allow it to find those dimensions where the data can 12:31.940 --> 12:33.400 actually be separated. 12:35.230 --> 12:40.840 Then because the community which we are using is actually being used to convert this data into higher 12:40.840 --> 12:41.480 dimensions. 12:41.740 --> 12:45.670 Now, when we already have higher dimensions, then it is easier task. 12:46.300 --> 12:53.230 Now it uses a subset of training points in the decision function called support vector. 12:53.250 --> 12:55.000 So it is a memory efficient. 12:55.570 --> 13:01.090 Now, again, same point because there is no need of all the points. 13:01.090 --> 13:03.190 All the data points are not needed. 13:03.190 --> 13:06.070 We only need the support vectors. 13:06.220 --> 13:09.240 That design will be more efficient in nature. 13:09.430 --> 13:16.420 And again, it will be better for me and it will not be as complex as other algorithms, which will 13:16.420 --> 13:21.790 take a lot of time to do to actually get trained. 13:23.660 --> 13:30.220 Now, what are the funds now, it does not perform well when we have large data set because it required 13:30.230 --> 13:31.730 training, time is higher. 13:32.660 --> 13:38.030 OK, so it will not perform well when we have large data said, because it will have to compare their 13:38.030 --> 13:39.970 distances and then find out. 13:40.130 --> 13:42.220 So it is better to have a smaller data set. 13:42.230 --> 13:45.670 We can clearly find the line of separation, then use it now. 13:45.680 --> 13:53.690 It also does not perform well when the data has some more noises like this that classes are overlapping. 13:55.240 --> 14:02.140 Now, here you can see when the glasses are separated, it is working fine, but when it gets closer 14:02.140 --> 14:05.760 to the data, it actually is not able to classify it. 14:06.580 --> 14:15.190 So when the glasses wave is overlapping, then it will not be able to find a line, a line of separation, 14:15.370 --> 14:18.040 and then it will not find a line of separation. 14:18.040 --> 14:20.890 Then obviously it will not be able to classify it properly. 14:21.460 --> 14:27.930 Now, SVM does not directly provide a probability estimate, so we cannot find a probability of a point 14:27.940 --> 14:33.100 being in one particular glass because it can StreetLink give which glass it will belong to, but it 14:33.100 --> 14:39.940 will not give a probability because anyhow, it is calculating based on the distance from the from the 14:39.940 --> 14:41.320 line of separation rate. 14:41.500 --> 14:45.790 So it cannot really calculate the probability of it being in one glass. 14:46.540 --> 14:53.440 Now these are calculated using the expensive five fold cross-validation it is included in the related 14:53.800 --> 14:57.790 SVC method of the Vitan Cyclone Lababidi. 14:59.420 --> 15:07.510 So this is about SBM in the next session, we will have a look at the whole walkthrough of SBM on a 15:07.530 --> 15:13.450 tax data so that you will have a clearer picture of how this is implemented.