WEBVTT 00:01.290 --> 00:07.710 In this session, we will discuss about the first machine learning links, unsupervised learning algorithm, 00:08.010 --> 00:10.560 which is hierarchical clustering. 00:12.210 --> 00:20.040 Hierarchical clustering is one of the popular and easy to understand clustering techniques, and it 00:20.040 --> 00:22.830 is also known as agglomerated clustering Begbie. 00:24.040 --> 00:33.310 So in this particular technique, what we do is we initially find out the distance between each data 00:33.310 --> 00:38.290 point and each data point is considered as a cluster innocent. 00:39.640 --> 00:48.550 Now, at each iteration, the similar clusters merge with another cluster on the one cluster or clusters 00:48.550 --> 00:49.040 are formed. 00:49.450 --> 00:51.100 So what they will do is. 00:52.610 --> 00:55.100 We will simply have. 00:58.500 --> 01:00.660 Different data, datapoint. 01:01.960 --> 01:07.110 So each data point will be considered as a cluster actual. 01:08.490 --> 01:16.980 Then we will create a matrix which will have the distance of each data point with other data point. 01:18.260 --> 01:22.670 Then we will combine the nearest data datapoints. 01:23.810 --> 01:26.180 In one particular group. 01:27.290 --> 01:34.040 Then once we have these Meurice data points, they will again be combined to. 01:35.600 --> 01:43.310 Another group and so on, so it will keep on combining different groups, fill the required number of 01:43.310 --> 01:45.700 clusters which are required and reached. 01:46.280 --> 01:55.290 So what we do here is it creates a hierarchy of clusters and presents the hierarchy, then drugan cortically 01:55.340 --> 01:55.910 structure. 01:57.080 --> 02:04.670 Now, this method does not require the number of clusters to be specified at the beginning because in 02:04.670 --> 02:10.910 the beginning all the data points will be considered as individual clusters. 02:11.210 --> 02:18.790 So we will be starting the process and going to the point where there is only one cluster left. 02:20.060 --> 02:28.130 So all the planes will slowly and gradually be convoy combined into one single cluster, now we can 02:28.310 --> 02:36.560 subdivide or stop at any point of time when we want to have or when we actually achieve the cluster's. 02:38.080 --> 02:42.650 So the distance connectivity between the observations is measured. 02:42.940 --> 02:49.010 So we will find out the distance between each and every data point. 02:49.180 --> 02:55.580 So to find out each and every data points distance, we will need to have them skewed. 02:55.960 --> 03:03.430 We cannot have age to be ranging from zero to hundred and amount to be ranging in age, because that 03:03.430 --> 03:07.660 will be having a very large difference in the skin. 03:08.680 --> 03:16.150 When we have large difference in the scale, then the distance, which will be between the ages, will 03:16.150 --> 03:23.650 be diminished with due to the distance which would be present in the amount. 03:23.920 --> 03:25.090 So what will happen? 03:25.210 --> 03:27.780 Because age will have smaller distance. 03:27.970 --> 03:35.680 So all that age data, all the data points which are closer in age, will be grouped together and there 03:35.680 --> 03:37.600 will be no impact of. 03:38.840 --> 03:40.070 And the amount of. 03:41.480 --> 03:50.410 So what we do is in this, we apply a bottom up approach, so each observation starts in its own disaster 03:50.750 --> 03:55.640 and the similarity or the distance between each cluster is computing. 03:57.120 --> 04:05.390 And then we merged the two most similar ones, each operation, until there is only one cluster left. 04:06.410 --> 04:13.670 Now, these are the big computing, the distance between each and every cluster after every iteration. 04:14.900 --> 04:24.200 So you can see that this is so much computationally expensive because there will be a lot of competition 04:24.200 --> 04:25.680 which will be required for this. 04:27.660 --> 04:34.890 And because that is such a high in need of computation required, because of this particular reason, 04:35.190 --> 04:43.940 it is feasible to use agglomerated flustering or hierarchical clustering only when we have a small dataset, 04:44.610 --> 04:46.570 if they have a large dataset. 04:46.590 --> 04:49.830 It will take a lot of time to compute this. 04:51.490 --> 04:58.240 Now, a three that shows how clusters are merged, split or split hierarchically. 04:58.570 --> 05:00.550 This is what the underground is. 05:01.530 --> 05:08.280 Each node on the tree is a cluster that is each subtype here is a cluster. 05:08.640 --> 05:13.760 If we go further, deeper into it, there are some clusters internally present. 05:14.340 --> 05:19.590 So let us see if we want to have two clusters, then we can simply divide it here. 05:19.590 --> 05:23.880 And one cluster would be this one and another cluster would be this one. 05:24.940 --> 05:31.750 In case I want to have three clusters, then I can divide it here and now I have three clusters. 05:33.360 --> 05:37.410 Similarly, if I want to have four clusters, I can divide it here. 05:38.320 --> 05:42.040 And then I will have one, two, three. 05:43.090 --> 05:44.580 I would clusters in her. 05:44.890 --> 05:47.440 So this is how we find out the clusters. 05:50.560 --> 05:54.510 Now, let's talk about a few properties of this clustering technique. 05:55.570 --> 06:05.310 The first one is that it is computationally heavy, that is if we have a small dataset, then it will 06:05.320 --> 06:06.580 work really fine. 06:07.000 --> 06:14.130 But if we have a very large dataset, then we will not recommend someone to use hierarchical clustering. 06:15.280 --> 06:19.010 Next is gives same cluster every time. 06:19.240 --> 06:24.610 So we have several different clustering algorithms which will give different clusters. 06:24.850 --> 06:32.010 But hierarchical clustering is one method which will always give the same clusters every day. 06:33.170 --> 06:40.070 So there is no need to provide the number of clusters at the beginning because it will create the clusters 06:40.070 --> 06:47.120 in the same method I at the time or after the clusters have been created, we can decide what number 06:47.120 --> 06:48.110 of clusters we need. 06:49.250 --> 06:55.800 Lastly, it is having a difficulty in handling different sized clusters and the irregular shapes. 06:56.060 --> 07:05.090 So if there is data where we have clusters with different size or irregular shape, then it will be 07:05.090 --> 07:06.800 a little difficult to handle. 07:08.710 --> 07:14.450 So it will, because it will always try to combine clusters in a field of two. 07:14.590 --> 07:19.840 So that is why it will always kind of end up having clusters of similar sites. 07:20.620 --> 07:25.660 So this is about hierarchical clustering, how we actually implement this. 07:25.870 --> 07:29.110 We will have a look at this in the next session. 07:29.390 --> 07:33.270 Then we will implement this particular code for hierarchical clustering.