Types of Machine Learning: Supervised, Unsupervised, Reinforcement

Key Takeaways

hello everyone welcome to the session so I hope that uh before watching this video you already have covered up my previous session where we have talked about introductory of machine learning what is machine learning all about right uh then we have talked about the history and evolution of machine learning and we have seen that uh this domain is not something which is coming up now it's been 1980s time where the concept of machine learning was coined right now uh in this session I'll talk about the types of machine learning that we have because uh considering you all as a beginner uh we should know that how many different types of data that we have followed by how many categories of machine learning algorithms we have which we will cover up in the upcoming sessions okay so without wasting any time let's get started to understand the types of machine learning so here whenever I'm talking about the types of machine learning I will categorize into you can say um three components usually you will see in maximum standard books and websites that categorize into three components but I'll talk about the fourth component as well in this session okay uh so basically the very first version which you will see is a you can say supervised machine learning algorithm supervised machine learning algorithm this this is the first category of a machine learning algorithms so let's talk about this in a great detail before moving to the next category okay let's consider as we have seen that whenever we are talking about the concepts of machine learning algorithms we are trying to take historical data and we are trying to train that data set right this is what we have understood in the previous session now let's say that I am providing you one problem and U that problem is belonging to a healthcare domain and specifically if I'll talk about the problem statement let's say I will provide you the data set of patients which are having diabetic and which are not having diabetic so there are people who are diabetic and there are people who are non diabetic listen my problem statement very carefully okay when I'm saying diabetic means it belongs to class one when I'm saying non diabetic means it belongs to class zero okay now what is the problem statement obviously uh you have to design a model which gives me a prediction right which gives me the output whether it is one or zero means whether the patient will be diabetic or not now obviously to determine that whether the patient is a diabetic patient or not we require the uh input data you can say we required the records of the patients which are diabetic and which are non-diabetic so that our model will try to learn the patterns this is how basically the algorithm work right so what we have is let's say I'm providing you set of features for example we will be having feature fub1 FS2 F3 F4 F5 and corresponding to every feature when I'm saying feature it means the columns that we have we will be having having a Target value now this target value is either zero or it will be one zero indicates a person is non-diabetic one indicates a person is diabetic this is a numeric data that you have and you can say that maybe they're providing a BP of a patient insulin level of a patient like that there are five features that we have and corresponding to every feature we will be having a Target value which is either zero or one means either the patient will be diabetic or the patient will be non diabetic now basically for every record for example if let's say you will be having one to 1,000 number of Records okay for that you have a Target value which is labeled as either zero or one and then you are trying to provide a training to the model splitting your data into training and testing how the complete flow will go uh I'll explain in the later part of the sessions also but as of now just think over it that we will be having a kind of a data set this is something which we call as a data set composed of rows and columns where the columns are something which we called as a features so we will be having columns which is what I will say in my sessions as features of that data we will be having rows which I will say as a set of records that we have or data points that we have in machine learning terminology this is what we us ually say now just try to understand what I'm trying to convey here for every particular row or for every particular record that I have I have a corresponding Target value which is either zero or one right what I'm saying is I have a corresponding Target value which is either zero or one this kind of problem where you will be having for every record the target value the label value is something which we call as super vised machine learning algorithms wherever you will be having a data which is having a Target value available to you so to train that kind of a data you will always have a supervised machine learning algorithms so what is the meaning of supervised machine learning it means that simple in literal terms it means that labeled or you can say Target data is given in the data set in the data set simple very very simple I hope it makes sense to everyone what I'm trying to convey here Target data or you can say the label data is given in the data set that is something which you can call it as a supervised machine learning now in this case specifically where we have to provide either the value as zero or one is something which you can call it as a classification problem classification problem okay so always remember one thing that the first thing is according to the given data you have to determine that whether this is a supervised machine learning or not if it is supervised how will you determine a Target column is given to you or not if it is given then it's supervised now you have to check that in that data what is the target you want to uh you know predict either it is zero or one or it is something else what is that something else maybe continuous values for example let me give you second problem statement so this is the problem statement one where we have talked about the first example which is of healthcare domain so this is the example number one right example number one now I will provide you example number two again again try to listen my problem statement very carefully let's say I'm providing you again the set of features feature F1 feature F2 feature F3 feature F4 and feature F5 and I'm saying feature again I'm saying it's a column and corresponding to every feature you will be having a Target value and in this case Target value is the house price house price and you want to train your model in such a way that after learning from this data set your model will be able to give me the prediction that if I'll provide you the features F1 F2 F3 F4 F5 what is the optimal Target value of a house price let's say the feature can be the number of B bedrooms the number of bedrooms right is it a 2bhk or 3bhk or 4 bhk depending upon the price level will be varied then locality or the area right then maybe uh number of bathrooms and so on so these are the features corresponding to these features our uh price of the house will be varied for sure now here if you will observe the price of the house is something which is a continuous value maybe you can say it is a 2.04 CR you can say uh the first record house price is this second corod house price is let's say 1 .23 CR third grow uh third row house price is let's say 50 uh lakh and so on so basically what is happening here if you will observe again if I'll ask you is this a supervised machine learning you will say yes why is yes because the target value is given to you corresponding to every record so again let's say if you will be having 1,000 records so corresponding to those 1,000 records if you will observe a Target value is given to you a Target value is given to you in the complete 1,000 recuits so here if you will observe it's a supervised one but it's not a classification problem problem per se this kind of problem is something which we called as a regression problem regression problem I hope it is clear to everyone so what I have just now explained I have explained you about what is a supervised machine learning where I have said you that wherever in a given data set a Target value or Target column is given to you that is something which we call as supervised machine learning and in that supervised machine learning we have categorized into two components one is something which we call as a regression second one is something that we called as a classification classification right right regression meaning the target value that you want to find out belongs to what belongs to continuous numbers okay you want to predict some continuous numbers classification means that here the target value belongs to some class there will be some class either 0 1 2 it can be binary classification it can multiclass classification problem so let's go back to the screen and what we have seen in supervised machine learning we we have seen that we will be provided the label data and it will be useful on classification as well as on regression task so this is the important point that you have to understand make sense now let's talk about the second category that we have which is unsupervised machine learning so let's go back to the screen now whenever I'm talking about this unsupervised machine learning this is the second category so until so far I have talked about first category now let's talk about the second category which is unsupervised machine learning what is this all about let's try to understand this concept as well when I'm talking about this unsupervised machine learning it is just the opposite of what we have discussed just now for the supervised one meaning let's say you will be having a data set but in that data Target value is not available so in any kind of a data where Target column Target column which you also call it as a labeled column labeled column if it is not available if it is not available then that kind of problem is something which we call as unsupervised machine learning problem for example just a very simple example which I I want to give to you is uh I hope you went to a grocery shop maybe a big Bazar in the malls or the Hypermarket where you will observe that uh let's say people are working there okay and there is a different different blocks are there for every particular particular uh product that we have for example grocery items are at one block uh then you know um maybe sary items are at another block maybe uh you will observe that Maggie and all those stuffs are at other block you will observe that all the body wash are at the other block so what you will observe is that there is a different different blocks that we have where the products are being assigned if let's say tomorrow a new person will come in that particular Hypermarket and uh he wants to work in that Hypermarket what do you think is there any uh problem he will be facing to assign the products no right because there is a different different groups that we have so he will just check it out which particular group uh that body wash body wash or Maggie belongs to and we will go and directly put it there so what I'm trying to convey here please try to understand wherever you want to uh classify the customers or classify the uh things which you want to work on that's where the concept of clustering came into picture clustering came into picture so clustering is something that we have in unsupervised machine learning for example I will say to you that you know um classify the credit card ratings or on the basis of a Cil score on the basis of of a credit card uh score just tell me with just tell me three groups three clusters one group who's having a salary of maybe more than one LPA another group which is having a salary between 50 uh th000 50,000 to let's say 1 LPA 1 LPA right or having category Which is less than 50k right which is bigger than 50k so there is a three customer base we have C1 C2 and C3 on the basis of the features that I'm providing to you so what I'm providing is the details of the financial transactions obviously banks will provide that and if you will observe uh you will get a loan and you will get a lot many calls which is very frating at times uh on the basis of what on the basis of your uh cill score on the basis of your transactions that you made every month they usually call and provide a loan to the customers and that loan will always be helpful uh for the uh Bank people because that's where they will generate the revenue if you will not able to repay the amount so by saying that on the basis of the patterns of the features that we have we can Define the Clusters we can Define the similar groups we can Define the similar groups for that we don't have the target value as of now but we are trying to create that so that's what we called as unsupervised machine learning where the concepts of clustering algorithms came into picture so if I'll just go back again to the screen here we are training on unlabeled data that is what I have explained useful in clustering and anomaly detection I'll talk about an anomal detection as well it's kind of outliers that we have in a given data means extremely uh awkward points that we have in a data set I'll talk about that in a later part but yeah this is what we have called as unsupervised machine learning so until so far we have talked about two categories one is the supervised one another one is unsupervised one how will you get to know by given the data set you will be able to understand if the target value is given to you it's a supervised one if it if it is not given to you it is a unsupervised one very simple right now there is a third category which is not mentioned in the screen but I'll talk about that which we call as semi-supervised semi-supervised machine learning you might uh use this particular thing in the industry a lot now what usually happens is that in a real world practical scenarios we don't have very huge amount of label data specifically in various domains like healthcare so what usually happens in in that scenario let's say uh we will take a example of a healthcare domain itself because in healthcare labeling is very tedious task because doctors will not have the time to label and without doctors we can't as a normal human being label the things as diabetic or non-diabetic cancerous or non-cancerous depending upon the uh given features right only doctors can do the labeling and they don't have the time to be very Frank so labeling is a very tedious task now what usually happens let's say we will be having a problem of again I'm taking a same example diabetic or non-diabetic means it's a classification problem where we will be having a class of one and the class of zero let's consider a scenario now let's say I'm saying that out of thousand records out of thousand records that we have 500 records are there which are labeled which are labeled means we we will be having a Target value corresponding to that okay but there are remaining 500 records that we have we have a patient data of 500 records but we don't have the there is no target value which is available in that case now please try to understand this is a different situation from the situation that I have talked about in the above cases whether it be supervised or it be unsupervised in supervised we completely have a th000 records having a Target value that's what I have explained in unsupervised we completely have a set of thousand records which are not having Target value but now I'm talking about semi-supervised where maybe almost half of the records are having a Target value half of them are not having Target value and that's the kind of a situation that you will feel in a real time industry use cases so when you will enter in the industry or whenever you will do a research work uh in any of the domains you will find out that there are specific domains like I have given you the example of a healthcare where this is a very biggest issue we will be having a half of them are supervised half of them are unsupervised then what what we will do that's the biggest question that kind of a category is something which you can classify it as a semi-supervised because half of the data is supervised half of them is not now the question is that how we will be able to solve this problem right so the question is how we will be able to solve this problem now what we can do in this case what you will do you will say that what I will do is first of all I'll focus on this 500 records which are not having any Target value and this is what we usually call it as a unsupervised one right because just now we have discussed unsupervised machine learning are those which is not having a Target value so what you will do you will apply the clustering based algorithms on this data clustering algorithm now what will happen when you apply the clustering algorithm you will be able to get two clusters one cluster belongs to class zero one cluster belongs to class one so you will be having cluster Z and cluster one it means that now for your 500 records where the labeling was not there you have made a label maybe 250 records are there almost which belongs to class zero and 250 records are there which belongs to class one now you got the labels now what you will do you will merge these 500 records which were not which were not having Target value but now we have because of the clustering algorithms we merge them and then now we will be having in total Again th000 Records 500 records are already labeled 500 we have labeled by a clustering algorithm again we have th000 records and now what we can do is we can apply the supervised machine learning technology G algorithms that we will discuss in the upcoming sessions what are the algorithms that we have but now our problem of giving the prediction of zero or one will be solved now you can ask me one more question that why is there is a need to you know add these 500 records why can't we directly go and do the task completely on these 500 records the answer is we can do but higher the data set that you have there is a high chances that you will be able to get a good accuracy It's Kind the better knowledge you will be able to gain the more knowledge you will be G you will be able to gain the more expertise you will be able to have for that subject right similarly when you are providing uh you are providing the data set make sure that you will provide enough data so that your model will be able to give you the accurate predictions right so if I will be having 500 records patients data that that is so crucial why will not I utilize that so I will utilize that by doing the unsupervised machine learning algorithm first getting the predictions first clustering clusters first then merge them with the remaining 500 recuits that's the simple scenario which we usually use in the industries as well now let's move to the fourth category which was the last one which is the reinforcement learning so usually the third category which I have talked about just now usually people not considered as a separate one but just to make you guys uh clearly explain the this the concepts I have made it separated okay reinforcement learning so when we talking about reinforcement learning this is different altogether from what we have discussed in supervised semi-supervised and unsupervised what it says is that it says that I'll train the agent or the model and it will do the action on the basis of the environmental factors environmental factors or environment variables right what is the meaning of that I hope that you must have seen self-driving cars or you must have at least heard of the self-driving cars right many people have also seen that at least in tier one uh cities not that much as famous but yeah uh we have outside India you will be definitely able to see in us and uh many uh good developed uh you know cities you will be able to see this self driving C now the self-driving cars is completely built upon this reinforcement learning where on the basis of the environment the model will take some action right what is the meaning of environment for example let's consider the scenario of a self-driving car itself let's say if any obstacle will come automatically that self-driving car will apply a break so on the basis of that environment it will take some action now when the training usually happens for these cars either a reward W will be awarded or a penalty will be given depending upon the action so if that let's say it's called as an agent if that agent or the model will take a good action will provide the reward which indicates that okay you are taking a correct action if it is taking a wrong action we'll provide the penalty that is how better the action higher the reward uh uh bad action means hire the penalty this is how the complete learning usually happen happens depending upon the re environmental variables so if I'll go back to the screen you can see that here the training on agent will be happening to take the actions that maximize a reward so here what we want is ultimately we want to maximize that reward means we want actions to be very much AP because if you will talk about self-driving cars it's a it's a very important use case where if suppose a person is completely relying on the car uh we have to make sure that that car is working properly fine wherever there is a red signal is there it should stop wherever there's a obstacle is there it should stop right uh wherever they have to change the gear it should change like that so that is something which is a part of reinforcement learning so this is the overall talk about today's session where we have talked about specifically the types of machine learning we started with a supervised machine learning where I have talked about that given a data if suppose a label data is given to you it's called as a supervised one I have given you a practical use case of a diabetic patient data further I have classified that supervis into two important problems one is regression another one is classification the second important category which I have talked about is unsupervised machine learning where I have said you that if there is no label data is given to you that's something which we call as a unsupervised one where the clustering usually happens means similar items will be grouped together the last one is something which we have talked about reinforcement before that I have talked about semi-supervised where I have again provide you a use case of a diabetic patient where half of the data is a labeled one half of them is a unlabeled one at the very end we have had a great discussion about reinforcement where we have talked about the example of a self-driving car and have tried to you know understand the use case where depending upon the environmental variables agent will take a action that will maximize the reward so this is the overall idea behind the types of machine learning in the upcoming section I'll talk about now the applications of machine learning so stay tuned and I'll see you all very soon in the upcoming videos