Steps for creating a Random Forest So, you have your entire dataset, and you have divided this dataset into subsets. We have the same number of rows, but the rows are not the same, right? So, we don't have the same data in different sets. We then train n number of trees, providing each subset to each individual tree. We have randomly selected columns for each split on a tree. Once that is done, we know the result of every tree. We then take the voting for every data point, and wherever we are getting the majority of votes, that's the class which we assign. So, that's at a higher level. How we create a random forest: This is, you know, now an example of creating these subsets of data. This process is called **bootstrap** (or bootstrapping), the creation of subsets of data. For example, with the same variables, we have now four different rows. What we'll do is, we'll create subsets, and every subset will have four rows only, but we'll pick them at random. Some of the subsets would have repetition on different rows. so let's see how it looks for example this is one of the data sets okay on the left hand side i have the original data uh and on the right hand side i have the subset so if you see we have four rows in the subset as well however this row here and this row here are repeating so this row 75 has been taken twice which means now some some tree will take these two rows only and repeat them uh both some tree might take first and fourth row and repeat the fourth row thrice so this kind of permutation and combination will occur and you will have n number of subsets so this is what is called bootstrapping of your data Now, once you have your bootstrap dataset, so this is the same dataset, right? This is the same subset with two identical rows. This time, we are training a decision tree on this dataset, but we are not taking all the columns. So, here we are only taking these two columns, and we hide the other two columns from your decision tree, meaning these two columns are not shown. Now, this node here will split your data based on these two columns only. The next tree can get a different set of columns, and that process will continue to happen until... so, as it says, instead of all the four variables, we are going to take only two variables. This is your bootstrap dataset. So, we are taking the bootstrap dataset using a random subset of the overall columns and then building a decision tree. Again, on the next node, we can take a subset, which, you know, suppose this was partitioned on high BB. On the next node, we'll hide this column from this next, you know, node. We will offer these three columns, and as I mentioned in the beginning, based on different algorithms like, you know, Gini index, you can read about these when you study decision trees, or entropy or variance reduction. So, automatically, it will pick which is the best column out of the columns which you are exposing it to, which will make your data pure. But the idea is that we are letting different nodes take different columns to do the splitting of your data. That's why most of your decision trees will be different from each other. That is the important concept to understand. So, eventually, what will happen is your original data would be used through different datasets to create multiple decision trees, and all the decision trees will be different from each other. You keep on doing this until you have a significant number of trees. For example, here it says ideally you should repeat this step hundreds of times because we are talking about a real scenario where you have a lot of columns and a lot of rows. In real practice, real projects, you would see there are, you know, more than 100 variables and hundreds of thousands of rows. This kind of algorithm will make a lot of sense, but then you have to create a lot of decision trees. When I say you don’t have to do all of this manually, right, you simply take a function, and your function will do all of this for you. But you need to know what is going on behind the scenes, what is happening behind the scenes, and then you have to change the hyperparameters. You can only do that once you understand all of these. You know, details—the last thing is we are now going to take the votes. So, this is the dataset, which was the same example we looked at, which was repeating. Now, what we’ll do is, we’ll take this data row and check each and every tree that received this row in the dataset, and we’ll see which class that tree has predicted. So, if out of 100 trees, 95 have predicted this as "yes" and 5 have predicted this as "no," we’ll take this as "yes." Now, you might ask me, "What if there are 50-50 or what if 60-40? How do we know which to take?" But, if you think intuitively, this dataset is, you know, one of the rows from the dataset. So, that kind of situation will never occur. Most of your trees will be predicting it in the right class only. Intuitively, if you think about this, there is a significant relationship between these attributes and this class only. There will be a few trees that have used a specific set of columns, causing them to make a wrong prediction. But this voting actually is the thing which gives you very high accuracy when we use random forests. So, you will never get into a situation where you will get 50-50 votes for most of your dataset. If that is the case, what will you do? You will change the number of trees that you're building. Even though that never happens, but if you continue to face this problem, you know you are not picking the right algorithm.