 Hello everyone, welcome to the session of Introduction to Machine Learning. Today, we will discuss the basic overview of machine learning, the basic concept of machine learning and what are the different features that people study as a part of machine learning techniques that also we will study and where it is used specifically in the business domain, couple of examples also we will discuss. As a part of data science, machine learning is one of the most integral part of predictive modeling or say AI process. So, the first question would be what is machine learning? Machine learning is a subset of AI process which is very important in data science field especially for the prediction purpose and here this model learn the data behavior, the past data and the data pattern and it is understand the underlined behavior and predict the future accordingly. It train the data and accordingly it define a best model, suitable model through algorithm and then iterative algorithm process. But let us see the basic one of the important aspects like it focuses on algorithm that learn and improve from the past data and mimicking the human learning and increase the accuracy over a period of time because the more data comes into the system the machine behave better based on the data pattern and predict effectively. So, the more data new data comes to the system it is becomes more accurate based on the prediction model. Here I have you know given a general layout in a in a graphical picture, we can see data science is a overall aspects and AI is a much superset of machine learning and there are many other aspects like you know data differentiation making or say data analytics which we have been discussing throughout many sessions of predictive analytics of business forecasting where we have studied supervised and unsupervised learning. Today also I will give some couple of aspects of supervised and unsupervised learning as a part of this machine learning session, but you will get to know over a period of time in this session that what the what is the core of machine learning and why it is so important in industry specifically for the data scientist like you know aspirants. So, why machine learning? Here you can see couple of points that I have mentioned it unlocks the hidden insights from massive data sets and going beyond the traditional method. It automates the complex tasks and free us to think for more statistical aspects to involve yourself for more a statistical aspects and more and complexity are been addressed through the machine learning aspects as well. Machine learn the intelligent models and create in the data pattern and learn the trained of the data and create a intelligent model. It enables the data differentiation making we have discussed detail in the first week about the data differentiation making and we have covered this basic aspects of machine learning there too. Here I am repeating that aspects also like you know it adapts and improve with the new data handle ever changing situation because this new situation comes once the new data comes suppose if you go to the say grocery stores and initially if you think about couple of data the buying pattern of the behavior or retail store you might say that you know that these customers will buy this type of product and what kind of products will be sold in ABC type like ABC classification type that you can understand in the initial data. But once you gather more data for a last couple of months probably your accuracy level level of ABC classification or which customer will buy what type of product that would be done in a better manner through machine learning models of the techniques are supervised or the classification models or say clustering methods. So they are the advantage of machine learning the more data comes the more the model become accurate as well as the it gives a better prediction models. It uncover the hidden pattern driving innovation and advancement across industry. I will show you couple of applications area where you can see that why people are using desperately this machine learning techniques in different fields of their application domain. Now the next question would be what is the basic difference between the traditional programming and machine learning. Remember we do C programming C++ Java you know Python is a which is a part of your platform right. But overall if you think about the overall basic programming what we learn is that it is actually the traditional process of defining some quote or some different model right and that you quote and you get the outcome there it is a control is there in your hand but overall where it comes to the you know the explicit instruction are being given to the algorithm or to the programming that you do this way. But when it comes to the machine learning the machine learn the data behavior and it improves the pattern of the data and accordingly it makes a prediction. We do not force machine that you follow this particular we do not give instruction machine learn based on the data pattern the obviously we do guide that you know what could be whether supervised or unsupervised or what type of supervised method you learn to follow that we provide. But overall it is a based on the pattern it learn the data pattern ensemble method and then you know to some extent we call it as a machine learn the data pattern and accordingly the based on the trained algorithm it is predicts the future. Look at here in basic traditional methods you have a clear control of your model right and the tasks that you are going to do that you know. But in machine learning it is a adaptability machine learns and improves from the new data and handle the changing situation look at the different aspects. Here it is a fast development because you know quickly you can do it because you know the model and but in machine learning you have a complex task. It handles the complex situation over a period of time and it is manage the difficulty much better than our traditional programming algorithm. But there in machine learning what happens is that the strength is that over a period of time it learns and though it takes time. But once it is done it becomes a very solid prediction model as per as you know data science are concerned or you know future business aspects are concerned. So they have many advantages. I am not denying the traditional models of programming techniques are not bad. But this is coming a superset when it comes to the machine learning. It is learn the data pattern and behavior and like you think about you know say reinforcement learning. They are you know machine totally learn everything based on the data pattern and accordingly they adapt the changes the model and they come up with the best prediction based on their data nature. So that is the second like upgraded part of machine learning we are not focusing on the reinforcement learning. But basic understanding of machine learning we should focus today. Look at the interpretability in the strength you know as per as traditional programming you can see it has a good interpret like easy to understand the logic behind the program. But when it comes to the machine learning you can think about you know it makes the data different predictions without expressive programming. So it has a better prediction power but interpret is concerned this is easy. But here making the interpretation of your model to some extent complex but over a period of time it is becomes a structure algorithm right. Here you can see the traditional graphical difference of traditional programming and the machine learning programming. You can see here it is a data and the program that you fixed and accordingly machine give the output based on your model or algorithm. But when it comes to the machine learning algorithm your data level data are being given as input imagine your level data whether the loan will be disbursed or not or the person will die or survive or you know you can think any example whether the image will be recognized or not based on the past data you have a level you have the features along with the data that output data information will become as the input of your system and the machine will learn the data pattern and accordingly it will come up with the algorithm and corresponding predictions. Here you can see one example here suppose customer buying behavior you have the data and the machine learn the pattern based on the data and some input data based on the once the structure is ready you can think about new data with the training and testing which I will discuss today and then you know your model will learn the data behavior and it will come up with the predictions whether the customer will remain there loyal customer or not or is a fault customers or who is buying most grocery products the women are buying more grocery product or men are buying where they are cancelling more all these things you can do based on the machine learning predictions. So, these are basic difference but the next question will be where to apply in is there any application domain where people are using machine learning? Yes. In today's world in each and every sector people are using machine learning when you open your mobile and you saw some movie in YouTube or say you know you recognize something some photo etcetera machine understand that this fellow is interested or this person is interested about this type of movie this type of cricketer this type of an artist this type of books this type of novels. So, they accordingly recommend next time when you open this you know YouTube or say you know your Google news feed similar type of product comes it is a machine learning right machine learns your interest and accordingly its predicts. So, here you can see couple of examples that I mentioned like facial recognition technology unlocking the smartphone identifying individuals photo security and personalization. So, there are many applications you know recommendation system it is self driven car you know different type of music generation through machine learning techniques you can you know generate different type of music tunes and you know the fort predictions here are couple of points I have mentioned, but here you can see couple of you know industry oriented applications I have mentioned like you know stock market trading you can you apply the fort predictions you know visual personal assistant spam whether the mail is a spam or the you know genuine mail that filtering process can also be done through machine learning based on data pattern like hello good like greetings kind of what it comes in the headline and the subject line people might sometimes think that the system might sometimes think that it could be a spam email. So, based on the data pattern that different type of mail comes the system understand the machine understand which mail is spam which mail is a genuine mail. So, accordingly it categorize and it is post put the email into different boxes. So, self driven cars which are the most application area of machine learning people are using it right. So, product recommendation tariff prediction speech recognition image recognition automatic language transactions or you know translation language translation are been one of the most important application of machine learning remember when you open any document in say Hindi or in different language regional language what you do you translate it right and then you see there also people use and you read as per your own preferred language or the English or you know your particular regional language and you read it quickly right. So, it is all this a part of machine understand and accordingly there are many applications like these days people are talking about generative A.I etc. all these things, but it starts from the basic data science data analytics and then machine learning and the deep learning and then it goes to the A.I process and generative A.I and the bigger data science superset one by one you know the enhanced knowledge one by one medical diagnosis everywhere you will find the applications of machine learning. These are the you know basic understanding that why it is so important and where you can apply. But today since I told about that it will be a basic overview session on machine learning I will give the basic couple of features of machine learning and couple of different type of machine learning technique and few four three four topic popular topic I will discuss through examples. Let us see first the common terminologies used in machine learning understand this couple of terminologies which are very important which will be helpful to you to read when you go to the deeper syllabus of machine learning and if you take a detailed course of machine learning you need the basic understanding of machine learning terminologies right. So, this you know couple of slides you will understand the basic terminologies when you enter into the specific model in detail. For example, first is the data definitely it is the source it is the well right of your model. So, of your training system. So, you need the raw material for any production planning right or any system to manufacturing system to produce. Similarly, here also you need the data which will be a your well to train the machine to train the model. But that data can be structured in a tabular format or you know column wise effectively given or it can be unstructured also like the extreme age or you know different type of comment people use in different facebook or you know you know social media or say twitter. So, they are different unstructured data there you can categorize them you can classify them is the techniques of machine learning and AI and then you can classify that. So, first point is the data source right authentic data source which you required as the initial input to your machine learning model. From there your machine learning techniques will start. But another question is that once you get the data it is a machine that learn based on the data right and they train based on the developed algorithm based on couple of techniques that you provide or it is given to the system of machine learning based on that that machine or the computer will understand the data pattern right. For that what you have to do they need some initial data we call that as a training data. Suppose you have a list of data say this much of data. So, this say 75 percent will keep as a training data and then based on that you develop or the model will be developed right model will understand and the pattern will be whether it is supervised unsupervised or whether you want to do a classification or clustering or random forest whatever you want to do that first will be done or a decision tree first will be done based on the say 75 percent of training data. So, once the model is developed then you need to do the like you know testing that developed model or algorithm will be now tested based on this test data maybe 25 percent you can keep for your suppose or say 80 percent you can keep here and 20 percent you can keep for testing purpose or 90 percent 10 percent depending on your choice and the decision making or the management requirement. So, this is called two aspects of data division one is the training data another is the test data. Now, the next point is that validation set what is the validation set of your data it is actually you know there are many way people define this analysis you take the raw data into two part first say training data and next say testing data right test part that I told about say 75 percent and say 75 percent. Now, among the training data sometimes what do people do you know among the training data sometimes people do divide into two part one this is again your training and this it is a validation first you validate your model this also you can divide into two part once that is done like you know develop your model with say another percentage of data say another 75 percent of it you can take another 20 percent you keep as a validation part. So, once you develop the model and it validated whether it is a doing a effectively or not and then you modify your model let model learn and fine tune the best algorithm best fitting suitable of predictions and then once the little bit of inter validation are done then they go for the testing whether the model is really working for accuracy level the perfection level like you know for the 95 percent accuracy or different accuracy level measurement are there we have discussed a lot right. So, that comes as a part of final testing. So, the validation is a intermediate testing data which will be segregated from the training data itself. So, there are three components you understand one is there is a training, training and then testing and then again training can be divided into two part may initial training and the validation and then testing. So, this way you know this three component of data classification are done in the initial level training testing and validation. Here you can see it is a representative of subset of entire data sets used to train the model for internal parameters and enable it to learn the pattern from the data. Then what is the testing model unseen data sometimes people call it is unseen data also or the next remaining that remaining that this 25 percent unseen data that objects evaluate the model real world performance the accuracy level the performance whether the in insurance sector if you want to apply machine learning then in that case it is a accuracy level. So, how the model is working but what is the validation the training data the training data is classified into again fine tuned in setting and prevent the overfitting of the unseen data what is overfitting and all these things I will discuss in the next slides. But let us see once you understand the three aspects of data division training testing or say training validation and then testing then this comes the features what is the features of the data features of the data are nothing but suppose you have a data sets and suppose it is the ID of the data and they say it is a you know level data and then couple of features features one like age different type whether man or man and the you know kind of the academic career on all these things different data sets are like there are different you know features are there or you know columns are there. So, all this called as a features characteristics based on that you define your level whether the customer is a loyal customer or not. So, level is a specific you know output cell or you can say the final performance measurement of your model which you are not touching you are to some extent you are using it as input data to train the model but that is your final performance measurement right. But there are many other components you can say the input data sets or you know unlabeled data we call it as a characteristics of the features data based on these features we define our model the train our model the machine learning algorithm comes to finally see the performance of your level data the part see the performance of your level data say the second one say level data. So, this is what you know feature featuring like there is another topic called feature engineering we are not going to want that part this is the characteristics of the features of the data describe data. Then algorithm which you develop the model develop a specific set of instruction used to train the machine learning model different algorithms are suit for different tasks right based on the situation where sometimes you use say linear regression sometimes you would say logistic regression sometimes you use decision tree sometimes you random for a sometimes you use k-meansal clustering so different support vector machine random for the different techniques you know principal component analysis you use based on different suitability and the data requirement right. Then model a representative of the learning knowledge from the data which is actually the final version of your outcome of your model that this is the my structure this is the my decision support system you use it for your company or for your data sets. So, this is what overall initial you know terminologies of machine learning then we will go to the next level of understanding of terminologies common terminologies the level part which I have already discussed level part is the most important part of a data sets of machine learning. Suppose this is your level and these the these are the characteristics features right these are the features whether you will buy or sell the stocks or you will buy or leave the store these are you know your level this is the final decision recommendation right the person has heart disease or not remember the examples that we have discussed in different sessions. So, these are the level but there will be features right different characteristics will be there. So, this based on this characteristics data sets we will decide the level as output or say decision making performance. So, that is called level a category or the value associated with the data point used in supervised learning tasks that is the supervised learning that you are trying to supervised ok and level provides the ground truth for the model to learn from remember this is the ground truth based on that your final inter prediction or the performance of your model will work this is also one of the part of your data sets, but we are defining that as a level right that would be like you know when you define a regression say y equals to say mx plus c your c and y x and y both are data sets right, but here y is your output data right supervised data that you are predicting that is your main value. So, here also we consider this y as one of the level sometimes we define as a value sometimes we define is a yes no binary kind of decision making. So, that can be defined in a different manner of supervised learning right like if you go to logistic regression then it is y is a value, but if you go to logistic regression this will become a binary value yes or no kind of thing y will not consider the direct value you will consider some problem. In the next session we will discuss detail of logistic regression you will get to know more detail about it. Now prediction prediction is the outcome the forecasting value from your trained model. Evaluation the performance measurement of your model right based on the unseen data test data not the training data this is the test data this is your test data and this model or you have developed the model you have developed that is from the training data or say training plus validation data look at that I have discussed in the previous slide. Now the three more component that is called overfitting or three characteristics overfitting underfitting and cross validation very important terminology in machine learning overfitting or nothing, but suppose you have the two data set right say training data and you have a test data right. So, two type of data you have classified the data say 75 percent or 80 percent here and 20 say 80 percent here and 20 percent you have kept for the trained testing part for you know final whether the model is performing based or not. Now this training data the 80 percent training data you have and you have based on that you have developed a learning model algorithm model and you have developed your structure right. There in case you are based on the training data if your model perform well, but when you use that model the machine learning algorithm with the test data your model does not perform well. So, in that case we call it as overfitting look at here a phenomenon where a machine learning model become overly focused on the specific details and noise presence within the training data this leads to a model performing with the training data very well, but for the trained testing data it does not provide efficiency effectively, but for training data it looks very good, but poor but become very poor performance when it is used the test data or unseen data underfitting means whatever the algorithm you have developed or the model you have trained or the outcome has come not suitable at all not for even training data not for even testing data this type of algorithm fail to develop a structure therefore we call it is a failed machine learning model or say it is a the process is failed the algorithm is failed. So, we call it underfitting model then the cross validation this is very interesting. Remember I told about the data into two part training and testing right sometimes what happens now or there are another two component called you know training sub training you can say and the validation right validation. So, when you divide the data into two part say you know training and the validation you have divided this say 90 percent from the first and the 10 percent from the last right as a testing or say validation. Now, what you do you take change the data pattern like do random sampling. Suppose now you take the first 10 percent as your testing and the remaining 90 percent as your training. And then you again you take the from the middle. So, sometimes you do this random sampling and you do the analysis of training testing and validation. And what happens over a period of time your model becomes very like a ensemble model model become very strong over a period of time and it becomes a generalized structure of your machine learning algorithm. So, this is called cross validation the validation part are up and down not only from the initial data. It is like you know you want to say study say sample of data which party will win for the upcoming say general election. In that case what you do you take a data sample that you have collected from the people directly say. And then you divide who are could be your say you know training data and who could be the testing data. And then you take come up with a prediction and then you validate it. Then again you change the sample and then you do put a different set for the testing different 10 percent or say 20 percent for the testing. Then you do again the prediction and then again change the data this way this enter process of iteration are called the cross validation. A statistical techniques to use to evaluate the model performance mitigate the overfitting part that we have discussed overfitting term which model is most suitable because sometimes in overfitting model training model provide a good outcome but testing data does not fit the suitability of the model. In that case this process this cross validation helps you in overfitting the situation. The data is divided into multiple folds with each fold uses the training and validation in term in each fold in each iteration this provides a more robust and generalized estimation of the model performance. I believe it is clear to you. Now, once you understand the basic terminology let us come to the classification the types of machine learning problem. Here I have mentioned the two types which are most popular which are common in machine learning models or methods. One is called the unsupervised learning and the second term is called the classification or called the supervised learning. In unsupervised learning there are many methods. Remember in unsupervised learning you do not have dependent variable. You do not have a leveled variable. All features data are there based on that you classify the data. Based on that you will classify the data who are male who are female or who buy most who do not buy. So, they are the classification you done clustering process you do. Through factor analysis through support vector machine linear discriminant analysis through singular value decomposition and there are different dimension detection methods are there association rules are there text mining process are there. So, all this falls under clustering process grouping process. Then the another part of machine learning is the supervised learning where you have the dependent variable we have the level variable. There you train the model and you make a prediction right. In that case there will be a classification method like logistic regression decision tree random forest today I will give the basic definitions with examples of each of them. And then regression models like linear simple linear regression multiple linear regression we have discussed detail of all of them right. Then forecasting model like moving average exponential smoothing Arima SCF PSCF we have discussed all of them in our previous sessions. So, all these part models falls under supervised learning. You can the bagging boosting are also you know very popular terminology in supervised models where you know you divide the sample into different bags and then use your different techniques of boosting process and the you know overall random forest or say you know whatever the technique or say decision tree you want to use. So, for that you do the bagging process to you know to create a better ensemble models and support vector machine neural network deep learning all this can also fall under supervised learning. So, that is the bagging boosting process.