WEBVTT 00:01.470 --> 00:06.960 The next model, which we will be discussing about is K nearest neighbors. 00:08.420 --> 00:13.590 But the name you can understand that this is about nearest neighbors. 00:14.420 --> 00:23.810 So here we will be considering the data point and from the point we will find out the nearest neighbors 00:23.810 --> 00:25.730 of that particular data point. 00:27.580 --> 00:35.500 And based on the nearest neighbors, we will decide what to us should the particular point should be 00:35.500 --> 00:43.870 assigned to or what value should the value of the particular data point be getting. 00:44.320 --> 00:48.320 So what value should be predicted out of that particular point? 00:49.150 --> 00:50.010 So let us go. 00:53.780 --> 01:02.410 So here is the number of neighbors, so let's say we have this particular data and we have assigned 01:02.750 --> 01:03.480 to be three. 01:04.370 --> 01:10.820 So in this particular data as gays three and we want to find out the class for this particular study, 01:11.150 --> 01:20.450 then we will consider the majority point that this this three, hence this red star is actually belonging 01:20.450 --> 01:21.970 to these four people. 01:22.190 --> 01:24.980 Plus that is Class B. 01:26.330 --> 01:34.970 Now, if you consider me as six, then it will fall into this particular region, that is the nearest 01:34.970 --> 01:35.940 six volumes. 01:36.680 --> 01:45.440 Now, when we are considering the nearest six in this scenario, the majority vote is of these yellow 01:45.440 --> 01:47.390 classes that this class. 01:48.680 --> 01:57.200 So now we can see that how the nearest the number of the key value decides what should be the value 01:57.200 --> 02:00.420 of the point, which we are looking for. 02:01.290 --> 02:09.800 Now, the small value of K passes over everything that is due to learning the localized five. 02:10.340 --> 02:14.690 It is not able to find the general bargain. 02:15.470 --> 02:22.720 So that is the reason why, because this point is looking for a very small number of key. 02:23.030 --> 02:30.170 That is why it is not able to identify that this point is actually belonging to Class E. 02:32.080 --> 02:39.700 Next is when we have a very large give and then it will become highly generalized. 02:40.390 --> 02:43.240 I mean, Mr. Complex bought it. 02:43.510 --> 02:50.750 So let's say that the nearest key value is actually a very high value system. 02:51.610 --> 03:01.840 So in that case, the nearest ten point would be one, two, three, four, five, six, seven, eight, 03:02.410 --> 03:04.000 nine and 10. 03:04.750 --> 03:06.520 That is all of the values. 03:07.060 --> 03:12.960 And then we take a majority vote on the glasses actually have five five point. 03:13.510 --> 03:17.790 So we will not really be able to analyze what this actually means. 03:18.880 --> 03:27.260 Hence, we need to find out a value of care, which is not very large and not very small. 03:28.060 --> 03:33.140 So we have to make sure that the value of K is acceptable. 03:34.030 --> 03:41.800 So for this, what we usually do is we find out the results for different key values. 03:42.490 --> 03:44.870 We will be finding out the results. 03:44.890 --> 03:52.690 Work is equal to one, is equal to two, is equal to 10 or 20 based on the amount of data that we have. 03:54.020 --> 04:02.900 And then we will compare that which particular value gives the best result, and that is the key which 04:02.900 --> 04:04.160 we will be selecting. 04:08.810 --> 04:12.260 Now, let us look at a few properties of Ken. 04:13.290 --> 04:22.080 So human can be used for both classification as well as regression, predictive problems. 04:23.040 --> 04:30.840 Now, when you want to classify so in case of classification, we will take the nearest key and take 04:30.840 --> 04:36.900 a majority vote of the point around it and then decide which class to the point below. 04:38.970 --> 04:46.570 When we are working with regression in case of regression, we will be considering these points and 04:46.570 --> 04:49.710 taking a weighted average of the values. 04:52.640 --> 05:01.340 A weighted average will allow to regularise the values because some values might be close and some points 05:01.340 --> 05:02.640 can be fired up. 05:03.200 --> 05:10.640 So a weighted average or weighted average on the basis of distance will allow to modulate the value 05:10.640 --> 05:14.530 in case of finding out the regression values. 05:17.160 --> 05:21.420 It is mainly used for classification, predictive problems. 05:22.200 --> 05:25.510 Why is it mainly used for classification, predictive problems? 05:25.740 --> 05:32.220 It is mainly used for classification problems because it can consider that it is a distance based in 05:32.220 --> 05:39.520 Wertham and it might ignore a few buttons because it is looking for the nearest point. 05:39.690 --> 05:43.110 So it is looking in a circular region around. 05:45.950 --> 05:54.380 Now, it is called a lazy learning algorithm devised by the lazy learning algorithm because there is 05:54.380 --> 05:56.830 no specialized training phase. 05:58.350 --> 06:09.630 In this gordito, it uses all the data for training classification, so it will not need explicit training 06:09.990 --> 06:17.430 period, when we are actually making a prediction, we can simply provide all the data points and these 06:17.430 --> 06:18.480 data points. 06:18.480 --> 06:21.990 It says we act as the training. 06:21.990 --> 06:29.480 What if the removal of your data point, then it will be classified differently. 06:29.640 --> 06:34.680 And if the other data points, then it reclassified different. 06:35.370 --> 06:39.100 So that is why it is called the lazy learning called. 06:41.010 --> 06:48.870 Then it is known the Matrix learning algorithm, that is, it does not assume anything about the underlying 06:48.870 --> 06:49.170 data. 06:50.350 --> 06:59.770 And it finds the it's based on closely matching data points that it will consider the closest data point 07:00.040 --> 07:01.910 for this particular criteria. 07:01.930 --> 07:11.140 And because the closeness is measured by the distance, which is the reason why we will be scaling the 07:11.150 --> 07:15.670 data, so we will scale the data so that. 07:17.060 --> 07:26.660 The distance will be equated for different types of features, so because there could be a feature named 07:26.660 --> 07:34.520 age and they could be another feature amount now age will be ranging from, let's say, zero to hundred 07:35.060 --> 07:37.790 and the amount might range in lacson. 07:38.600 --> 07:41.860 So in that case, we need to have skinny. 07:45.220 --> 07:57.190 Now, let us discuss a few pros, so in order to uncomplicated and easy to apply in an atom is uncomplicated. 07:57.310 --> 07:59.360 It is not complicated in nature. 07:59.530 --> 08:01.110 It is very easy to apply. 08:01.130 --> 08:07.120 So you can simply see if we have the data points and we know the distances, then we can easily apply 08:07.120 --> 08:11.140 the algorithm by hand and find out what the point should be. 08:13.440 --> 08:21.600 Then there are only two metrics to provide the algorithm value and the distance metrics, so. 08:22.930 --> 08:30.040 All we need for this algorithm is the value of work and the distance between the points so that we can 08:30.040 --> 08:36.910 just decide on the basis of that how many points we need to consider and we find the nearest point and 08:36.910 --> 08:38.560 then make our decision. 08:40.960 --> 08:46.700 Next walks with any number of classes, not just binary classification. 08:47.080 --> 08:54.280 So in this case, we might have any number of classes and based on that, we can just find out the distance 08:54.520 --> 08:59.000 and the majority number of classes and then we can be able to go. 08:59.030 --> 09:03.100 We can simply use this algorithm for five classes or ten classes. 09:04.900 --> 09:13.330 Then it is easy to add new data to the bottom because it is a lazy loner, so it does not need any specific 09:13.330 --> 09:20.970 training so we can easily apply more data to bellbottom and then we will be able to analyze them according. 09:23.660 --> 09:31.040 Now, let us just discuss about a few points, so it is computationally expensive, right, because 09:31.040 --> 09:38.450 it will be finding out the distance between all points because it is a distance, basically, and it 09:38.450 --> 09:45.560 will have to find distance between all the points and then it will have to find the nearest distance 09:45.770 --> 09:50.870 of all points from this particular point, and then it will make their decision. 09:51.740 --> 09:55.060 That is why it is computationally expensive. 09:56.510 --> 10:04.130 Next is it is having a high memory storage requirement because it will have to save all the distances, 10:04.430 --> 10:10.720 that is why it will have a high memory storage, then it is hard to work with categorical features. 10:10.910 --> 10:13.800 So let's say we have certain categorical features. 10:13.820 --> 10:15.650 Now we have categories. 10:15.650 --> 10:18.580 We have values like age and amount. 10:18.890 --> 10:22.280 In that case, it is very easy to find out the distances. 10:22.550 --> 10:25.250 But let's say we have classes. 10:25.520 --> 10:29.060 One is gender and another class is. 10:29.270 --> 10:37.450 Let's say if someone is married or not, then there will be a very difficult to find out. 10:37.460 --> 10:39.950 What is the distance between the what? 10:43.830 --> 10:50.100 Next prediction is for a big number of features, that is for a big number of features, it will have 10:50.100 --> 10:54.370 to find out the distance based on all the features being considered. 10:54.600 --> 11:00.660 So the dimensions of the distance which will be found out, would be very high, which is why it will 11:00.660 --> 11:05.550 be a little difficult to work with these and also. 11:07.080 --> 11:15.720 The sensitivity towards the scaling of data is very important and also there might be a presence of 11:15.760 --> 11:19.650 irrelevant features which might tweak the distances. 11:20.070 --> 11:30.450 So let's say we have certain data points like is an amount and we do not scale them properly, then 11:30.780 --> 11:39.330 they might get biased towards the amount and we might not really consider the changes which are caused 11:39.330 --> 11:42.250 in the age of the person. 11:42.480 --> 11:46.640 So that is why it is very important to make sure that the data is. 11:51.610 --> 11:54.430 So this is about an. 11:55.850 --> 11:59.270 Next, we would learn how we can implement in. 12:00.560 --> 12:07.490 And always remember, Kevin is also always learning algorithm, they know what everybody seems to have 12:07.490 --> 12:10.150 learned are all supervised learning algorithm. 12:10.550 --> 12:17.030 That is another algorithm named Kamins, which is an unsupervised learning algorithm, which we will 12:17.030 --> 12:18.450 be learning very soon. 12:18.860 --> 12:26.490 So people usually get confused between the nearest neighbors and the key means algorithm. 12:26.660 --> 12:34.760 So we will be discussing the comparison and we will be visiting again and again while we will be discussing 12:34.760 --> 12:39.850 gaming's so that there is no confusion between both of the constant.