WEBVTT 00:01.050 --> 00:08.100 How in this session, we will discuss about the next unsupervised learning algorithm that the clustering 00:08.100 --> 00:10.500 algorithm named SCAN. 00:23.110 --> 00:29.340 Whatever drawbacks we have in case of humans would not be best in case of Busca. 00:31.000 --> 00:33.160 So let us discuss about the scam. 00:33.340 --> 00:46.050 So in case of gaming's clustering, we needed the number of clusters and we used to create very good 00:46.210 --> 00:48.370 clusters in case of gaming's clusters. 00:48.580 --> 00:55.230 And we would fix two separate points based on which the clusters would be created. 00:55.600 --> 01:01.680 Now, because of those Troyes, the clusters would meet radically initial. 01:03.240 --> 01:13.320 But in case of the rescan, we will not be declaring how many cluster's people to babies can or although 01:13.620 --> 01:19.540 they take the number of clusters, so for this they become important. 01:19.770 --> 01:28.590 We will need to barometer's the first family there is Epsilon and which is the neighborhood size. 01:29.100 --> 01:35.910 And the second parameter is the minimum number of point, which is required to form a cluster. 01:37.080 --> 01:44.910 Now, in case of a baby scan, it will also detect the number of clusters and it also does not make 01:44.910 --> 01:53.940 assumptions of those vertical clusters, which was happening in case of gaming's in babies can it can 01:53.940 --> 02:02.160 be used for anomaly detection and it does not get impacted by the outliers because the clusters will 02:02.160 --> 02:07.380 be formed based on the minimum number of points, which we will be declaring. 02:08.420 --> 02:16.120 So if there are any outliers, then based on these minimum points, we can actually figure that out, 02:16.430 --> 02:21.440 those outliers, and not consider them for those clustered. 02:22.880 --> 02:29.880 So for this, we will go to the same website where we can actually visualize the clustering. 02:31.850 --> 02:35.840 So this is the website for visualizing the clustering. 02:36.230 --> 02:42.030 So here we can simply select the type of data which people want. 02:42.350 --> 02:44.770 So let's try the smiley face again. 02:46.790 --> 02:48.650 Now in this smiley face. 02:49.890 --> 02:57.210 We will have to provide the absolute value and the minimum point's value, so minimum point's value 02:57.420 --> 03:03.430 will actually define how many points are required to actually form clusters. 03:04.140 --> 03:07.470 So for this, the minimum point is stated. 03:07.470 --> 03:15.870 As for the absolute value is given as one, which means that how many neba we want to have, the higher 03:15.870 --> 03:23.520 the value of Epsilon, the more points will be considered in case of one creating the neighborhood. 03:23.700 --> 03:25.560 So let's run this. 03:31.640 --> 03:38.150 Now, I want to begin this, it will randomly start to the starting point and start creating the clusters. 04:26.340 --> 04:34.040 Here you can see how easily the scan was able to find out unflawed photo of the clusters which will 04:34.050 --> 04:34.500 present. 04:36.560 --> 04:46.010 Now, let us try to look at more details, that is try to run another visualisations so we go to restart 04:46.010 --> 04:49.190 and this time we select the Pembridge Smiley. 04:51.160 --> 04:54.590 And let's decrease the Absolon value. 04:54.970 --> 04:57.790 So now the Absolon value is zero point six zero. 05:53.610 --> 05:57.580 So here you can see it is currently checking for different point. 05:57.870 --> 06:03.750 So it started checking these points, which are presently in the middle, but because it was not able 06:03.750 --> 06:11.670 to find four point one particular place, so it could not really create clusters of of these points. 06:12.090 --> 06:15.270 So let me increase the epsilon value for this one. 06:20.170 --> 06:26.500 Now, you can see that because the absolute value was low, smaller neighborhood clusters were created 06:26.500 --> 06:28.930 because there were more number of clusters created. 06:29.290 --> 06:34.900 Now, if I will increase the value to let us say it for Epsilon. 06:38.160 --> 06:39.570 Now, let's run this. 06:42.220 --> 06:48.640 And let us keep the minimum number of points for clustered as three instead of. 07:06.010 --> 07:11.080 With higher value of epsilon, you can see the cluster making process is also Plaistow. 07:14.660 --> 07:23.000 And you will see that it is checking into each and every point for the cluster creation, but wherever 07:23.000 --> 07:27.800 it will find at least three point, there will be it will create a cluster. 07:45.240 --> 07:52.970 But you can see it has detected all of the point as outliers, so let us try one more scenario. 07:55.710 --> 08:03.180 Now, let us have a look at the exemptions, so as we saw individualisation similarly here, you can 08:03.180 --> 08:09.680 see that it is able to take the clusters, even though they are not spherical in nature. 08:09.870 --> 08:12.840 It will detect the clusters in whatever shape. 08:12.840 --> 08:18.120 They are fresh and it can detect different patterns which are present in the data. 08:18.330 --> 08:21.690 And it will not try to create vertical clusters. 08:21.690 --> 08:25.200 And it will also isolate the outliers. 08:25.200 --> 08:31.920 So it will keep the outliers aside so that we will know that these are actually not a part of the ditzen. 08:33.730 --> 08:41.620 So here is a comparison of Baby Scammon Kamins, so baby scan is able to detect these dooring separately 08:41.620 --> 08:45.070 like means try to create very few clusters. 08:45.340 --> 08:48.820 Same thing applied for these two harpoons. 08:49.270 --> 08:54.530 And here again, a baby scan was able to find both the different classes. 08:54.550 --> 08:55.720 So this is different. 08:56.680 --> 08:58.570 And so they're different clusters. 08:58.870 --> 09:03.170 Was able to group these together while here in gaming's. 09:03.190 --> 09:06.790 It just right through a cluster and spherical vs. 09:08.660 --> 09:15.080 Here you can see that it has created three different clusters in Wisconsin canings, which are spherical 09:15.080 --> 09:15.580 clusters. 09:15.590 --> 09:19.140 So it has evolved as good as K means algorithm. 09:19.700 --> 09:24.830 I'm here you can see the most important example, which is uniform distribution. 09:25.070 --> 09:32.330 So the scan does not impose clusters if there are no different clusters in the data. 09:32.360 --> 09:41.060 So here there are actually no clusters and this is able to detect that thing while gaming's tries to 09:41.060 --> 09:45.210 create clusters, no matter the president or not. 09:48.720 --> 09:56.140 Now, to think the quality of the posters which have created we have this scoring metric which is scored 09:56.220 --> 10:00.980 for you, it is also called Sillett Index. 10:01.470 --> 10:09.270 If the index value is high, it means that the object is very much the way it's all cluster and poorly 10:09.270 --> 10:11.000 matched to the neighboring cluster. 10:11.370 --> 10:19.560 So Senate coefficient is calculated using the mean distance and the mean nearest cluster of distance. 10:20.130 --> 10:30.850 That is a and so those select corporations will be defined as s.a given by by minus E divided by a maximum 10:30.870 --> 10:32.200 of NBA. 10:34.560 --> 10:42.770 So here is the average dissimilarity of I object to all of the objects in the same cluster. 10:42.960 --> 10:53.490 So it compares the inverse cluster of cluster distance V VI compares the distance of a particular weight 10:53.490 --> 10:56.880 with the of all objects in the closest cluster. 10:57.510 --> 11:00.830 So this is what's called a gift. 11:01.080 --> 11:04.380 So we want this sentence to be closer to one. 11:04.620 --> 11:10.140 If the school of school is close to one, it means that the clusters are well formed. 11:10.410 --> 11:17.520 And if the center of school is minus one, it means that the clusters are actually placed in different 11:17.820 --> 11:18.470 clusters. 11:18.480 --> 11:25.620 So instead of being placed in the cluster, a point has been missed class of misplaced in a different 11:25.620 --> 11:26.190 cluster. 11:26.550 --> 11:36.040 And when the Szilard index value is zero, it means that the clusters are kind of overlapping in nature. 11:37.500 --> 11:40.260 So this is about the range of the value. 11:40.260 --> 11:44.550 That is, if citified is close to one, the sample is well clustered. 11:44.730 --> 11:50.520 If civic value is zero, could be assigned to one of the cluster closest to it. 11:50.550 --> 11:57.630 I'm topolice equally far away from both the clusters that this means that it indicates overlapping clusters 11:57.870 --> 12:00.420 and absolute value is minus one. 12:00.720 --> 12:04.250 Then it means that the sample is misclassified. 12:04.500 --> 12:14.070 So one is the perfect score and we want the solid score to be as close to one as possible. 12:14.790 --> 12:17.190 So this is about six score. 12:17.490 --> 12:26.580 Next, we will implement the babies can include them and so that you will get a picture and be able 12:26.580 --> 12:34.200 to compare different clustering algorithms that have had a good game means and they'll be scanned. 12:34.350 --> 12:43.830 And as we already told, a DV scan is used when we don't have any spherical clusters and ketamine's 12:43.830 --> 12:49.590 works better than we have spherical clusters and we actually know the number of clusters. 12:49.740 --> 12:56.220 In case we don't know what is the number of clusters, then we can easily use the rescan by providing 12:56.220 --> 12:59.160 the epsilon and minimum number of point value. 12:59.820 --> 13:06.980 The scam is used when we have certain outliers and then we want to isolate the outliers and I want to 13:06.990 --> 13:09.120 consider the other players in this condition. 13:10.760 --> 13:17.810 And regarding how to head up clustering, so hierarchical clustering would be helpful when we have a 13:17.810 --> 13:19.040 smaller dataset. 13:19.280 --> 13:24.020 So this is one restriction which hierarchical data clustering has. 13:24.230 --> 13:30.950 Otherwise hierarchical clustering is very nice clustering method because it does not change the type 13:30.950 --> 13:35.000 of cluster with each and every one like gaming's. 13:35.000 --> 13:40.400 I'm the rescan, so it gives the same cluster everytime had the clustering. 13:41.360 --> 13:46.360 So this is of one difference between these three clustering and qualities. 13:46.610 --> 13:49.240 So next we will see the implementation. 13:49.430 --> 13:51.770 So you will get a clearer picture of the same.