WEBVTT 00:01.150 --> 00:07.690 In this session, we will discuss about the next unsupervised learning algorithm, which is key means 00:07.690 --> 00:08.230 clustering. 00:09.300 --> 00:17.760 Now, when I have explained the key and then Legatum to do you, I strictly said that this is one little 00:17.760 --> 00:21.960 guarded them, which is often confused with the key means. 00:21.970 --> 00:32.520 And so now we will have to go through again and means together so that we can actually draw a line between 00:32.520 --> 00:33.790 both of these called. 00:35.090 --> 00:41.930 So the very first thing which you need to know is what is key means so that we can draw a line between 00:41.930 --> 00:46.960 key means and can and so that you don't get confused between them. 00:50.170 --> 00:57.420 So what is key means it is just the names look similar, but the concept is a lot different. 01:00.120 --> 01:07.930 So in case of K means clustering, he is the number of clusters which we will be creating. 01:08.520 --> 01:19.890 So he means clustering, has a number of clusters, the mechanism that randomly initializes random centroid 01:20.160 --> 01:21.720 in the feature space. 01:22.970 --> 01:30.830 And at the time of initialization, the key points are not actual centroid of 30. 01:32.480 --> 01:38.330 So the data point are then assigned to the nearest key point. 01:39.370 --> 01:48.460 The same droids are then moved so that they are at the center of the current designated clusters, so 01:48.460 --> 01:59.020 the very first difference between K means and again is that the key in means stands for the number of 01:59.020 --> 01:59.800 clusters. 02:00.920 --> 02:09.340 Why the key in key nearest neighbors stands for the number of neighbors which we are looking for. 02:11.110 --> 02:20.410 Second and most important differences, that Cannon is in supervised learning algorithm, that is we 02:20.410 --> 02:32.410 use it for making predictions via the key means algorithm is used for classifying or clustering the 02:32.410 --> 02:32.980 data. 02:33.190 --> 02:37.390 That is, it is an unsupervised learning one called. 02:39.290 --> 02:40.310 Let us look for the. 02:41.230 --> 02:48.760 So here we have the data, so let us say we have these points, so out of these points, if we want 02:48.760 --> 02:55.180 to find two clusters, so what we will do is we will place two centroid points to it. 02:56.220 --> 03:04.800 Now, we have placed these tools and right these tools androids would have been present at any place, 03:05.730 --> 03:08.580 these are randomly placed in this particular. 03:10.590 --> 03:19.800 Now, once they have placed these Synthroid, they then assign these data points, the points which 03:19.800 --> 03:26.880 are nearest to this centroid, to this particular cluster, and the points which are nearest to the 03:26.880 --> 03:31.020 second centroid to a cluster which belongs to this particular. 03:33.270 --> 03:36.210 So now the boys have been assigned. 03:37.380 --> 03:45.570 So the Redpoint belong to the cluster with the radix as Android, and the blue points belong to the 03:45.570 --> 03:49.380 cluster with blue X as the centroid. 03:50.220 --> 03:52.480 Now if you look at the data. 03:52.710 --> 03:56.380 This doesn't really look like this in the auditioner. 03:58.020 --> 04:06.900 So now the centroid will actually move towards the center of the cluster, which these androids have 04:06.900 --> 04:07.540 created. 04:08.400 --> 04:14.420 So now the points are moved to the center of the clusters. 04:14.580 --> 04:21.600 So the Red Point is brought closer to the red cluster and the blue point is brought closer to the blue 04:21.600 --> 04:22.140 cluster. 04:23.550 --> 04:35.540 Now, once this has been done, now these points will again be rearranged so that the point assigned 04:35.540 --> 04:37.530 to blue cluster. 04:38.870 --> 04:46.910 Are having blue as the nearest point and the clairvoyants assigned to the red cluster are nearest to 04:46.910 --> 04:48.410 the Red Point. 04:49.070 --> 04:56.090 Now, again, once this has been done, the Red Point will be moved to the center of the data, which 04:56.090 --> 05:00.900 we have, and no point will we move towards the center of the blue data. 05:01.610 --> 05:09.590 And then again, the points will be reassigned based on their distance from the center right. 05:10.580 --> 05:19.430 This will keep on repeating until the points do not move again, until the centroid stop moving. 05:20.180 --> 05:22.010 So this is what he means. 05:22.010 --> 05:25.770 Algorithm is here, the things to notice. 05:26.000 --> 05:35.120 Number one, it will try to create circular clusters because it is looking for the area around the center. 05:36.200 --> 05:44.810 Second thing is that no matter what kind of distribution is, it will still try to create a cluster. 05:46.070 --> 05:54.830 Next point is that whenever the planes are assigned, whenever the same is pleased, so based on the 05:54.830 --> 06:01.910 placement of these Synthroid, there would be different types of clusters which could be generated in 06:01.910 --> 06:02.460 the end. 06:02.990 --> 06:07.640 So it is not a sure, sure thing that we will get these two clusters. 06:07.940 --> 06:15.060 The clusters might be based on the placement of the centroid at the initial point of time. 06:15.470 --> 06:23.930 So after every initialization, we might get different clusters of different brands of the good. 06:26.840 --> 06:34.610 Let us look at the assumptions which we have, the first assumption is that key means assumes the variance 06:34.610 --> 06:38.340 of the distribution of each attribute to be spherical. 06:38.690 --> 06:41.840 That is, it tries to create very good. 06:43.430 --> 06:53.210 Cluster's, second thing is that already you have the same mediant, that is all we will be having the 06:53.210 --> 07:01.060 same variants and here again, we will have to make sure that the data has been scaled properly. 07:01.070 --> 07:06.520 Otherwise, the shape of cluster's, which we will receive, might not be as expected. 07:07.820 --> 07:17.150 Next, the prior probability for all key clusters is the same that this each cluster has roughly equal 07:17.150 --> 07:18.490 number of observations. 07:18.890 --> 07:26.060 So whenever these clusters will be created, they will have roughly equal number of observations. 07:26.360 --> 07:28.020 And apart from that. 07:29.240 --> 07:40.160 Outliers will not be considered separate, outliers will also be included in these clusters, so a position 07:40.160 --> 07:45.830 or a placement of an outlier can actually impact creation of these clusters. 07:48.090 --> 07:56.400 So here you can see that this data is in uniform mixture and there is no actual cluster present in the 07:56.400 --> 08:05.910 data, but just because we have pleased to say androids and the androids had to make clusters and thus 08:06.030 --> 08:10.310 it has created two clusters in unknown cluster data. 08:12.380 --> 08:20.570 Next here, you can see that the clusters are expected to be of the same size, that is, the clusters 08:20.780 --> 08:25.850 will have similar number of data points in then. 08:26.120 --> 08:32.970 So here, because this is a concentrated this has and this has more scattered. 08:32.990 --> 08:37.420 That is why it looks like they have different sizes of data. 08:37.430 --> 08:42.230 But actually they have seen size of the present in the clusters. 08:45.180 --> 08:50.080 Next game means assigns to spherical fiesta. 08:50.100 --> 08:52.260 So here we have a very good cluster. 08:52.530 --> 08:54.000 So what will happen? 08:54.330 --> 08:58.140 We would have expected that it will create cluster. 08:58.380 --> 09:04.250 One cluster would be the outer ring and another cluster would be the inner area, inner circle area. 09:04.560 --> 09:12.630 But actually it has created spherical clusters, cluster this half to one cluster and the other half 09:12.630 --> 09:17.570 to the other cluster, which is not a good clustering technique. 09:19.930 --> 09:23.890 So what are the rules for this particular algorithm? 09:24.250 --> 09:26.050 Number one, it is simple. 09:26.530 --> 09:28.960 Number two, it is flexible in nature. 09:29.350 --> 09:32.630 Number three, it is suitable for a large dataset. 09:32.890 --> 09:36.670 And lastly, it detects this very good Lesters very well. 09:36.670 --> 09:43.440 It will be able to detect such clusters very nicely and these clusters, which are spherical in nature 09:43.450 --> 09:44.200 very nicely. 09:44.230 --> 09:46.630 So this one, again, it has detected very well. 09:46.960 --> 09:55.990 But where we have some data, which is not really in very good form or here, you can see that these 09:55.990 --> 09:57.820 are actually outliers. 09:57.820 --> 10:02.740 And there are only two cluster sectors, one, this cluster and this cluster. 10:02.950 --> 10:05.610 So they are actually only two clusters present. 10:05.980 --> 10:11.210 So it was not able to determine that because we have already provided the number of clusters that we 10:11.210 --> 10:11.380 need. 10:12.130 --> 10:18.580 So no matter if there are extra clusters, present or not, it was to create one extra cluster instead 10:18.580 --> 10:22.180 of creating just two clusters which are actually present in this data. 10:24.480 --> 10:32.340 So what are the cons that is sensitive to the initial Synthroid location, so based on the central location 10:32.340 --> 10:38.170 of this dog, the Thingo Spherical Cluster, then it is sensitive to outliers. 10:38.170 --> 10:42.470 So that is here we have these outlier values present. 10:42.780 --> 10:52.290 So it actually tried to include it in a cluster and hence ended up creating another cluster and adding 10:52.290 --> 10:53.490 them into a cluster of. 10:57.050 --> 11:01.920 Next is it always creates a spherical cluster, which we have seen here. 11:02.630 --> 11:04.820 And lastly, it detects the. 11:06.130 --> 11:14.230 It is not applicable for categorical data because a categorical data will not have so much difference 11:14.230 --> 11:18.550 in the distances, so we cannot really use it for a categorical data. 11:19.240 --> 11:24.520 Next, we will see a visualization of the K means clustering. 11:26.530 --> 11:32.470 So this is one website where we can actually visualize the K means clustering. 11:33.010 --> 11:34.420 So here. 11:35.390 --> 11:44.000 We can select the initial centroid, so let us select the farthest point to be the initial centroid 11:44.000 --> 11:48.850 here and the type of detail like the date of the O. 11:51.090 --> 11:52.330 Up in my. 11:53.630 --> 12:05.800 So here we expect the number of Kluster to be for where we have one to three and four clusters, so 12:05.810 --> 12:09.320 let us start let us absent right to it. 12:10.460 --> 12:17.270 On the centroid, one more centroid and another centroid, and let us start training this. 12:22.090 --> 12:23.380 So it will. 12:24.430 --> 12:26.880 Assign the points to these centroid. 12:28.070 --> 12:35.840 Now we will update the same Droits, we will update the central location to the center of the data. 12:37.490 --> 12:44.540 So it has updated the central location and based on the central location, now we will reassign the 12:44.540 --> 12:45.170 points. 12:46.750 --> 12:55.180 The boys have been reassigned to the nearest business, and right now we will again update this central 12:55.180 --> 12:55.720 location. 12:57.850 --> 13:00.170 Now, the central location has been updated. 13:00.190 --> 13:02.080 Now we will reassign the points. 13:03.140 --> 13:09.200 So if we keep on doing it, we'll just do this again and again until this point is actually the same, 13:09.200 --> 13:10.520 that all of these data points. 13:11.060 --> 13:19.310 So it has created four clusters, which was provided by us because we knew that there are four clusters, 13:19.520 --> 13:24.950 but because it wanted to create a spherical cluster, it was not able to determine. 13:26.100 --> 13:33.510 The location of the property, so let us have a look at it again, let us restart this. 13:33.990 --> 13:44.610 So now let us place the center randomly and this time we will have the device can bring us the data. 13:44.640 --> 13:46.800 So these are the different clusters. 13:46.810 --> 13:54.300 These are the data points, letters and centroid as one, two, one, three. 13:54.310 --> 13:55.890 Let us have four centroid. 13:58.160 --> 14:01.650 Or just have to be sentenced for this and let us go. 14:02.210 --> 14:06.920 So it has classified the points based on dissent right now. 14:06.920 --> 14:11.870 Again, we will update the central location so it will bring the same rights to the center of the. 14:14.480 --> 14:21.170 Now, again, we will reassign the points, so the points which have been misclassified like this, 14:21.170 --> 14:27.440 red wine, these red points will be made for the college classes. 14:27.710 --> 14:29.950 So it will reassign the points. 14:29.960 --> 14:33.020 Now, the red points have been drawn into green and blue. 14:33.770 --> 14:36.200 Now, again, we will update the central location. 14:38.460 --> 14:41.970 Now, again, when we reassign the points, these points will be corrected. 14:49.570 --> 14:58.720 So you can see it is always trying to create circular clusters and it will stop when we actually reach 14:58.870 --> 15:05.030 the center of the data and all the points, all the clusters have almost equal number of data points. 15:05.590 --> 15:07.690 So this is where it has stopped. 15:07.960 --> 15:13.420 And you can see that there are almost equal number of data points present in all the three clusters. 15:14.050 --> 15:19.130 So you can try different types of clusters here. 15:19.150 --> 15:24.970 You can try different types of mixtures and then see what actually comes out of. 15:29.290 --> 15:31.840 So let us try the Bembridge smiley. 15:33.360 --> 15:40.830 Now, look at this, Bembridge Smiley, there are these outliers present, so it will it has actually 15:40.830 --> 15:45.270 tried to assign these outliers also in the clusters. 15:45.270 --> 15:48.170 So this is what we don't really want to have here. 15:48.450 --> 15:55.700 And the creation of circular clusters has actually impacted the cluster formation. 15:55.860 --> 16:02.000 So we want to have another variant or some other algorithm which might perform better than this. 16:02.400 --> 16:04.230 So it will perform better. 16:05.220 --> 16:13.020 Canines will perform better if we have spherically clusters so we can use it more, use more appropriately, 16:13.020 --> 16:20.420 if we have spherically clusters and we have data points which are spherical in nature.