Model Parameters vs Hyperparameters - Techniques in ML Engineering #machinelearning

Key Takeaways

foreign to another session in the data series we are thrilled to be here with you this evening for a session full of action-packed learning I am Neelam tapa part of the data science team at analytics Vidya for those who have joined us for the first time a brief disc uh introduction to the data session the data is a series of webinar conducted by analytics Vidya and led by top industry experts it is a fun way to understand the concepts of data science from the leading players in the datatech domain and as the name suggests is uh it's just one hour dedicated to data we are hopeful that this session are going to be great source of enrichment and value adding for our community members now um on to our session today which is the model parameters versus Hyper parameters techniques in ml engineering in this data are a speaker will demonstrate the significance of parameter tuning and performance of the machine learning model which improves with a more acceptable choice of hyper parameter tuning and selection uh techniques in details he will also be showing simple use cases for your better understanding I hope you are excited to attend this data with us so uh before we kick things off and I hand it to our speaker a quick recap of the housekeeping items we are recording the session and the recording will be available on your YouTube channel the link you can find it in the chat section uh please use the Q a section for asking any questions you might have during the session and we will do our best to answer them as the data R progresses or towards the end we will be sharing a feedback poll towards the end uh you are requested to kind of fill that up before leaving the session now on to a speaker in this session of data we have Santa Baba pandyan with us Shanta Babu is currently working as AIA architect and delivery with cognizant he has a good experience of 20 plus years in Information Technology idea he has done his PG program in artificial intelligence and machine learning from great learning and University of Texas over to you Shanta Babu the virtual stage is all yours yeah thank you and tapa this wonderful evening um so so let me share my screen and please confirm like we're able to see the content okay right oh yes excellent excellent okay guys so I could see around 60 plus uh the the folks acting joint so thanks everyone so I'm not sure like where you guys are belongs to so it could be a morning uh afternoon or evening um so this is a great opportunity for me to just uh spam my time with you guys so the today's topic uh it's like uh it's a model parameters and Hyper parameters so try to understanding that what it's going to be and how we can we are going to use this specific thing uh when we're building the model okay so and the the major techniques which involved um in when I when we're doing like a model machine learning uh building the model um aspects so I'll try to cover as much as possible uh because the area is really um with vast area um I mean like so the the parameter is just a simple term but so believe me in that it's like a very huge impact when we're building the model so that is that's the reason actually I picked up the specific uh the topic today so without further delay so let's we jump into the uh the content so right so guys so let's try to understand what is tuning basically so the basically the tuning so hope you guys like this is analog um um radio so I think I know I'm not sure like how many of you are familiar with this orders like probably you can see this one or nowadays we have some digital uh FM videos right in in our house so probably there can be like the package that you used to see so when we are going to select a specific uh the channels so what we used to do so we used to tune that um the one the provider over here slider yeah so this is a provider actually they used to tune and we get the exact what we're looking for so we are not sure about like which which FM you want right so 98.5 99.5 107.5 something like that so there are multiple things in our um in general way so we used to tune that and we get the the exact um right so keep this specific item in your mind and then let's be focus on the rest of the item in our session so that you can understand better so here I'm giving a two bit Circle that is nothing but model parameter as well as the hyper parameter it's like looks like a something like a Greek and Latin it doesn't matter so because it is already like it's available in our uh when we're building any model so it comes along with that it doesn't matter like memory on that right and the same way for model parameters there's a very crispy and very easy to understand right and so before this one I think you guys are agree with me so you as I said earlier so every ml engineers and data scientists work and Ma there is no difference like in the role base so Case by case like I mean like industry specific or or organization specific that might be a role might be different but at the end of the day both uh the scientists and ml Engineers used to do but nowadays they are giving you some drastic um Light Between the engineers and the data scene but based on my understandings of both are not much difference so ml engineers and data scientists must understand the significance of hyper parametic tuning well while selecting the right machine learning or deep learning model to improve the performance of the model so if you ask me like hey I don't want to I don't want to care about like hyper parameter parameter whatever it is parameters like parameters different way I tell you that I will tell you that in the coming up upcoming slides so hyper parameter is a very important thing you may say I can go with whatever available uh that uh what is that the value I can go ahead and submit and I can go ahead with this hyper parameter means yes very well you can go ahead that doesn't matter but when you are applying into the real-time scenario you are um the stakeholders you have to give the answers to these stakeholders that is very important so the specific the ml engineering ml projects totally differ from um um totally different from your uh um the normal application development and deployment in the production because the ml is totally different the way is handling each delivery stage is totally different so and each and every stage we have a so many um written goals on that so in that the specific hyper parameter tuning is a very important net and pulse I would say if it is very loose so let's say for example if your bike or your vehicle then returns bullet round boots is not in good good condition so you cannot drive your the vehicle right it's the same way the hyper parameter is performing the very important role if it is even in this the kind of Ocean Way of the ml implementation the hyper parameters playing a bigger role here so with this condition and agreement let's move on to the Deep dive so as I mentioned it's like it's a model parameter and uh the hyper parameter is a big big circle I would say but when I compare with the model parameter hyper parameter is a very huge thing because there are a lot of things in it right and when it's going to be model parameter it's a very simple understanding so easy you can collapse that and Hyper parameter it's playing the major role let's see what is that right so when you start with anything that starts with small steps baby steps basically we should say right so that aspects Y is equal to e m x plus C I think hope you guys know what is MX plus C right so this is a very basic linear regression as and and even is a simple linear regression model we used to go ahead with like the Y is equal to maximum C and I am going one step back in your college days so you guys like I get a chance to work out with these kind of like solving that linear regression model by the time even when I doing my engineering at all so I never bother about like giving the data and I used to do and finding the uh finding the m and c and then I'm finding the Y value of y after finding my slope and coefficients right so the same way here so let's the few minutes we will check this one so what it is like then only we can understand what exactly the parameter and the hyper parameter look at guys here with respect to the machine learning language Y is equal to MX plus C so here it's m x plus b doesn't matter right so m is nothing but a slope and D B is nothing but your intercept so what we are doing in the mathematical way so we used to figure out like what is the value of M by giving this this Sigma of some formula I don't want to confuse you guys if you get both doesn't matter please pause sometime then then we'll move on to the next one so these are the value given in general way in mathematical way so we can find out x minus X bar and Y minus y bar x minus where the whole square and blah blah whatever the thing and trying to mean and etc etc and finally we'll figure out something like what is the m value at the same time you'll find out the C value whatever then what I will do I will find out like what is the other values of Y and other other value of like the value of y so this is the equation if I know the value of c and my B sorry m and b and then what are the value for the X so I can find out that way all right so this specific the flow you can understand how it is coming up so forgot about the actual interdiction that's fine but when you are doing your any linear regression when you are being in your in your college or schools so probably in stats or in mathematic engineering mathematics whatever it is so you come across the specific the formula and doing all those things so all these things happening inside your machine learning algorithm when you are implementing linear regression simple steps but we are not going to discuss that one but let me say what exactly the model parameter and Hyper parameter it's a simple answer is a configuration don't worry about that because actually if I six years back when I started my transition from the analytics into the advanced Analytics right it's Advanced analytics something about the machine learning AI everything so when I'm doing that six seven years back so by the time so really it was confusing with the model parameter hyper parameter why we are doing that right so it's a simple configuration putting the values different different values and then finding the outcome that's very important right so that is nothing but the the configuration of the value I mean that parameters inside your given equation you forget about the equation at all if you if you call the like regression linear regression or classifier model or random for uh k n whatever it is you can go ahead and you can get the output like you will get some output but there was a lot of parameters we used to say as a syntax files we used to say parameters so before in the normal language we used to call it built-in function right so in the built-in function we have a number of parameters we have to pass the value and we will get the output for example seeing concatenation something like that right for example let's see so the same way when you're talking about the ml right machine learning process so there are a lot of functions inbuilt function which is providing with python notice we're using python right so even are also providing that back to the python is the robust one now at the moment right so during that function when providing we have to changes the value of for example if I'm asking you guys write a simple simple function to add the two numbers what do you guys to do you'll function defining function a comma B and you are passing the value a might be a two B might be three and something a maybe a 10 and B might be 50. so we will get the values right so that's a simple function the same way if you were configuring your function you are defining the function you are calling the function and same time you are getting the output from that and you are enjoying the one but same time when we are dealing with the model then we are playing with a huge amount of data we have to configure our the model with the help of model parameter and Hyper parameter that's very important make it very simple and this is a very very important the understanding one hope you guys are familiar with this formula right and this is like end of when you when you are doing a regularization techniques and all in um in uh ml so we used to we use this specific uh Sigma y one y i minus MX plus m x i minus C the whole Square to the square plus Lambda and weight I don't I don't I don't want to go into DJ make sure that what is the model parameter model parameter this specific portion this is nothing but model parameter and this specific Lambda is nothing but a hyper parameter make it simple okay so let me give the indication that what is the role of model parameter what is the role of hyper parameter and how these things will impact your model hyper parameter or supplied as argument to the model algorithm during the initializing them as a key and value right it's nothing but a is equal to 5 b is equal to 10 something like that and the values are picked up by the data scientists or missionary who is building the model in the iqt mode that's very simple so now we are reaching almost where we are closing understanding what is hyper parameter hyper parameter and model connector so both are same doesn't matter right so just I'm mentioning that hyper parameter and you can make it that is hyper parameter or um model parameter simple okay let me move on now and try to understand what's the definition the quick understanding the model this is a simple linear regression slides I'm putting like the you're Drawing the Line so model parameter or configuration make clear on that variable that our internal to the model and the model learns from on it owns using a data set so make sure that when you are dealing with the model parameter that is going to be here always always internal right now what a data set what you are using So based on that how it is possible you can ask me please wait I will explain that and before that if I want to explain that in simple statement you may selecting like the font size colors underscore strike through whatever things in your uh when you're building the document or we're writing emails something like that so we used to specify your what you like our standard for my font or size so in the case you can select and you're sizing that and better way that is nothing but you are passing the parameter to the system internally so that based on that what happening like you are complete the board document it's getting clear and even it's like when you're making a presentation or a solution in the email so we will selecting the different font says colors and based on these standards whatever it is right so you are doing a configuration for your container the similar way the model parameter hyper parameter working the same way but the model parameter this is purely internal how it is purely internal I'm coming to that later I'm coming begin because I'll just go back so this Y X Y is equal to MX plus C I would say the A C so here I have my m and b these are like model parameter so I'm getting this with the help of all these data you got it so I'm getting the slope I'm getting the like to The Intercept with the available data so internally the system will calculating and it's helping me to put up this yz4 MX page formula and I'm getting the value right but what is hyper parameter you see that in so hyper parameters are the parameters that are explicitly defined as ml Engineers by the ml Engineers control the learning process of the node guys actually when you are building something when you are building we are training anything like for example we have to find out like probably if you guys that have a good background on the ml at least not good but at least the minimum medium level or practitioner level or initial level so probably you will have to do a lot of iteration of finding the model best model what is the best model coming up so what you used to do we will change the numbers and we'll try to understand that but specifically when you're dealing with the hyper parameter we never bother about uh these things by default what is available we'll start using that but that shouldn't be real time scenario that's very important right so one cannot know the exact the best value for the hyper parameter for a given problem statement so the best value can be determined by either rules of thumb or trial or error what is that that is starting but the hyper parameter space hyper parameter space is nothing but what it's going to be a the complete availability the permutation computation combination of the the whole overall thing yeah this is the one right so look at here this is a three-dimensional diagram I have n underscore estimators and mini underscore relief and minions will spread all right so in this case I can have a number of options my actual the mini split from 4 to 14 and my mini leave going to be like four total here and again by underscore estimators value might be like 200 to 2000 or 200 so I'm keep on putting the different combination of all this parameter and I'm finding my first one you don't bother about that we have a lot of uh functions and methods that we have so those things will help you finding out that which one is the best combination okay because when you run this script so it will give the thousands of lines of output so no need to going and digging that hey what is that which one is the best one so where I need to do that no no no issues on that so only the thing we you need to understand what is the bottom line or how the the basic fundamental is working right so the basic Mantra is here so if you understand the basic thing a is equal a plus b whole square is equal to what is a Formula a square plus b square plus 2 a b right simple thing so if you know that one whatever the value you give that if you're writing a program you build that that will give the answer the same way like Y is equal to X plus C and it's going to be simple linear or multilinear or whatever it is so if you know the basic stuff how it is mathematically so that's it is an actually for when we are building the model if you have a strong knowledge of the probability and statistics and even more a lot of other the Laplace transform reverse Laplace transform there are so many mathematical things we can incorporate and we can get the use of that if you know the basic thing we have a functions we call it out because the value we'll get the output so think about this one these are the lot of options we have so by using different combination of the values we can provide and we can get that right this is nothing but hyper parameter space that's very very important right so now we've gone through that what exactly the model parameter and Hyper parameter and let me give the high level things like the what differentiating that how it is parameter parameters model parameter right I mean I mean it's a model but it should be a model parameter that's internal to the model this is external to the model you'll see that one by one essential for making a prediction right because actually when you are putting that one internally it will calculate and give that one prediction like if you want to like tomorrow going to be rain or not so this specific model parameter help you guys but the same but this hyper parameter it's optimizing the model we have to optimize the model and ultimately that will support your best prediction for your model that's very important and same way specified and estimated while training the model as I showed you right the tabular column finding the finding the like horizontal slope and uh intercepts everything so this will be taken care by internally that is nothing but specifier estimated I would say it's estimated estimated while trailing the model so internally it will it will get the data from the data set and it will create the m and see everything so that you can get that so if you are familiar with integration model so that is a simple function called intercept so if you call that method I think like the it's not directly it's like comes around regression dot then intercept means that we give the uh output and same way for scope the same way so that will be estimated and this will be this this specific hyper parameter setting before the beginning of the training of the model so we are planning to give a lot of options first time going to be a five second time is going to be at 10 and third time is going to be a 15 I'm going to be 20 something like that so after some point of time what will happen so you can find out the output oh right so when I'm giving the value of like y5 10 15 20 25 something like that by the time you're able to understand all right I'm getting the different answers which one is the best one you can go ahead of based on your option you based on like the output and you can very well you can check with your smes so that's not only deciding by the machine learning changes machine learning this over liquidity You can predict because the person actually more than 12 years 15 years in this in the specific Mr industry so they can say that hey this is the model you're getting all these output this one is the best one so go ahead this one or else you should have the statistian oral SMB subject matter aspect experts inverting so they used to get they used to give you have to share that list of output and then they'll find out okay this is the one they'll run through that the list type of item and they'll pick the best one it will take time so that's the reason that the entire the road map and pipeline for the specific ml implementation will take time right so I would say like minimum uh eight months to one uh 15 months or 16 months more than a year because of building up the data collecting the data and it's building there are a lot of stages in that right basically I do not go inside that the different phases in machine learning so when we're dealing that so ultimately we have to put up that the out uh what is that we have to set the different value of paper parameter and we get the notebook and coming to the the fourth point learned on set by the model by itself right so we are not passing any value because actually if you are familiar with the simple linear regression model they are not going to bother about slope or intercept no we are not bothered simply we'll pass that data set after uh after giving to that model split make make sure that actually testing and training are only for training perspective when you're building a model but after that the things will be directly to it goes to the the all the data will go to the that training model so that the model will get the output that's fine and this specific hyper parameter it's setting by the manually manually in the sense actually they have a configuration again there are different in the real-time scenario they will do in different way passing the value to that file so that the manually the data will be like loaded keep on changing that manually in the sense actually always going and changeable during the simple uh when they're building that that's fine but once after everything's get collected what they will do the model is all not sustained and I'm telling you the model is not all stable one today the model is doing good tomorrow the model which give the bad day bad bad performance because because of the data so when the data is like slightly changes from today to tomorrow the data will be like it's gone differently because we can't get the proper output right so by the time we need to do the like the different configuration that's the reason actually we come up with the Mr Ops Team so the deploying the model and the changing that configuration it's like not frequently I'm telling it's going to be a day it will not change sometime actually that is there might be like depends on the data so depending on the data that will be changed okay so as I mentioned earlier model parameters always like the dependent data set and Hyper parameter it's a purely independent because you are manually setting the value um uh to your parameters and you are ascending that and same way and estimated why there are a lot of options we have the gradient descent and so many things that we have I don't go inside like we selected a lot of options algorithms optimization under that we have we can use that and here we have estimated Performance Tuning by the doing like we have a grid search and the other things that we have we'll see then and beside the performance of the Unseen data and this is specifically quality of the model actually if the quality of the model would be great then only the performance model would be like good so ultimately the hyper parameter is a superpower than the model parameter so that's very important thing so the best example in linear regression is a coefficient super vectors actually super effective things and A and N you can specify the needs and same if on hyper parameter is a learning read and K well in KN and so many things yeah we can see that one big one okay so I think somebody's putting the question what do you mean optimization model uh sir we can uh take it in the end of the session all right yeah good right right good say this is the mission learning life cycle guys so so what we used to do but generally we'll get the data and we'll split the data and we'll try to do the training the model and testing the model so model finalizing and then putting into that and will outcomes will compare that so generally we will split the data as I mentioned 84 80 20 75 that's fine that's fine this is the basic thing actually when you are learning machine learning and only this is one but in real time scene is totally different because once a model find like there is no point of sending the data again to the test but we can monitor that based on that we can again it will do in the back end all right so steps to perform the hyper parameter tuning the model selection is very important because what kind of algorithm you are going to select that is very important and review the parameters right what are the parameters we are going to select that's very important and selecting the searching and right searching method so like we have like how to search the data and everything and apply the cross validation approach translation cross validation approach is is bit little bit away from this this discussion and assess the score of the model that's very important because one step after the specific things what we have like uh um across cross validation approach and all we have a k-fold validation at all so that that's little bit away from this uh the police swapping so I'll try to cover like in the next section if possible Right these are the major thing so right fine so so far your discuss about uh the the entire Journey like what is the uh model parameter what is the hyper parameter how it is different where it's playing the role and how to identify that that's fine so now I'll come to this one little bit later because let me discuss this one the time to discuss the few iPad parameter and their influence on the Node that's very important so this is one is a very familiar one guys hope you guys are know very well right so X and Y actually used to pass the value for the data set right and independent dependent variable and here the classical is the random underscore state is equal to zero so we get the train and test set across the different execution right the random step the random state provides the seeds for the random number generated in order to stabilize the mode if you are not specifying this one what will happen the things will be like one time will get that different output and again when you're running the same thing you will get the different outputs so you want to stabilize the your balance your model so we have to use these specific uh uh what is that randomizes so Random underscore state is equal to set so we get the same set of uh set and uh trying set across the different execution but if you ask me like it it needs to be something like that yeah you can change that if you want but when you started practicing this one this one so better you go ahead with the zero sometimes you'll be with 101 it's like Case by case different so the random number if you omit this one system will generate this number and leaves the unpredictable behavior of the model as I mentioned earlier like if it is not mentioned anything one time will give this something some value second time will give you something like you keep on trying that so that you'll be frustrated on that okay good and next sample is a logic Logistics regression classified this is one of the very important thing actually this is inverse of Regulation strength so regulation techniques we have that's a different one so regulation techniques is nothing but to suppress the noise and to control your entire data set and on the ground so in that perspective so the logistic regression they are giving the C value the large location classifier and it's it's related to the Lambda probably you can remember the specific Lambda value in my earlier slide if you want icon this one okay so this Lambda is is derailed from the specific C1 so if if you ask me like where I go and see means all the functions is built in right so it's everything should be available inside that so so that actually we couldn't done we couldn't find like how it is internet doing but we used to understand that it's very important C is equal to one by Lambda right so that specific regulation parameter it's very important so that it is actually in the indirectly professional in a universally professional to the C so that the based on the value of the C that we keep on changing if you ask me like if you're asking me say that if I give the C value what will happen try this right so we have to try this basically we used to go with the best standards thousand because these are the very minimal thing but when you come across the ml Journey that's very very important that's very very important to understand like building the model some cases we cannot give it just like the numbers make sure that but we have to make sure that what are the hyper parameters we have to use in your when you are building the any model using algorithm that's very important right cool and again that is one of the best thing like the decision pre-class classifier so here we have uh what is the uh citrians and so what is that entropy and maximum depth of the dictionary and again the random state so these are something like mathematical related things basically this specific thing determines the maximum depth of the tree that's a very important that is very very important right so the same way for entropy and we have the other options as well I don't confuse us there are a lot of options we have so we can use that right uh Guinea I think that is another option so Guinea I guess so we can use that but if you ask if you can pass that value different value instead of entropy you can pass the guinea values you can get the output and you can change the maximum depth value 3 into 5 5 into 10 something like that but the maximum depth the based on the your the data set size right but in real time scenario it's going to be very huge so you you can classify something like that like you know right the digital structure classifieds uh kind of like uh uh what is that it comes again the random course way so it's it's a complete Retreat structure so that so that we you have to decide and which one you should use and which which one you know should not use that so the maximum depth you have to analyze and you have to put the value right it's not a standard one you have to use three four five symptom something like that so it depends on the data volume of the data and the possibilities of your independent dependent variable combination over there so you can use that Max content that's very important and liso regression this one actually I mentioned that Alpha value either you can give the alpha value is equal to like we have a regularization parameters as a point one by default they used to give and again this specific and user regression also like being we are regularizing your regression so try to understand that when we are building the model not only collecting the data and we're doing the Eda and finding that selects of future selections and applying the algorithm and finding this code now on top of that we have to the hyper Paramus tuning and then we have to do the regulation techniques and then we have to find out the outcome it could be right root means where or any other things we can find out so after that only we can stick on to that and remember that there won't be like one model one data set will always I will support one algorithm no it can be support more than one so in the case between more than one so you can bring the insible model and you can perform on that and we have a voting methodology there are so many things on that bagging system there enter these ancible techniques there are so many techniques of that so that probably you might know that one and same way here the principal component techniques are very important if if the Futures the future uh I mean like number of columns the independent variable is a very huge number then we used to go ahead with the pce model because we cannot have like 100 100 features with like we could build prediction is very really challenging so that when you're applying the principal model then that will be give like a number of principal components so it will make sure that like it will give the four four different uh components or five different components and again when you're applying that as principle component analysis so there may be like a combination of different values when you're passing the 10 then your 50 column right so that will give the five column precisely and then you can take that value and you can play around with that after that you can apply the all models and you can get the output and same time if it is going to be a two means the system will internally find out like which one is the best and give the two columns but very careful on that when you are doing the principal components it's not so easy and it's a very useful time consuming I understood but we have to be very careful on that so I am telling so we are going for each not only for the principle components when you're building any model that's very important that's very very important right so it helps to reduce the number of independent variable probably you know that's in indeed about like the PC model right and the next one it's going to be a hyper parameter what is that sorry this is a k n method in k n method okay this is going to be neighbors going to be 5 and P is equal to uh one value is uh power power parameter value is like a value and Metric is going to be like mini cos P so this is actually these are the simple things like uh equivalence distance and Manhattan distance and all actually intellectual calculator if you guys are familiar with all this is a mathematical thing you can understand the main the main the reason behind these specific items that's very important if you're providing p is equal to one means then the system will Taken like equivalent to that Manhattan distance way it will work if it is going to and based on that it will work on that so that's very very important thing here right so so far we have discussed about uh the different uh hyper parameter in quick quick way so still we have around 20 minutes I'll try to cover and we'll go for some simple demo because it's just in 60 minutes time so I need to cover entail questions I am I think I skipped one thing yeah um yeah so here the steps as I mentioned select the right model that's very important as I mentioned like it's not going to be selecting your own algorithm nodes we can select a multiple thing but one by one right and review the all the parameter list in the model so when you are using any algorithm so make sure that in circuit client act Library if you go and check that there are the first specific uh uh what is that so specific uh algorithm you might have a n number of parameters so we need to find out that which parameter we need to use that that's very very important and finding the method uh right method to switch hyper parameter and space that's very important thing and same way uh they are playing cross validation method and K folder structure so that like it will improve at the performance as well right that's very important thing so initially this is going to be like this so once you introducing or hyper parameter tuning the model will it will get like we want to understand like hyper parameter tuning and tuning that uh your your value and you will get the output again you can sign and test the data that's very important okay good right go here so hyper parameter influence in the models so if you take like a linear model what degree of polynomial features we should use so that will give that information so decision tree what is the maximum depth you can allow that's what I explained that and what is the minimum number of samples required at the leaf node and decision tree so that's a random Forest will we help that and how many trees we should include so that then for that should support like neural networks and how many layers we can use that and same time see how many layers for like a gradient descent and the all the stuff so these are the major influence for the hyper parameter uh for multiple different uh sample perspective given actually there are a standard um what is that tabular sheet is that so that is actually the reference number for each and every parameter for each um algorithm so probably like if you get a time um so you can just search and you can get that just I couldn't like get the time to include in this slide so but that specific things are giving a guidelines what is the default value why why we should go beyond that so that aspects actually we have so many things in it so we have to be careful about that and same time uh they have some guidelines so it shouldn't be more than that for example by default depth is equal to 5 or k-fold structure actually the key mostly that will go with the five means as a standard one so they will give the detail information saying that why you want to go to the file why it's not below that below that it's fine but they will give you like the portion Corners that's so so that you can understand why it is important okay so that is a very very important thing all right and we have so many options like hyper parameter automation techniques like the manual search we have and we have an atom search security search and so many things we have let's quickly show some demo it's already x0.1 okay I'm not going to execute because actually I I didn't check that like all the files are available okay so it's already executive one so hope you know this one I'm using like uh the big money and deputies are data set and I'm doing all this initial um Eda process right and here I'm giving my train etc weightage for this one so when I'm passing the different entropy and guinea and splitter right and best maximum depth 50 and so look at this one the 50 basically I'm giving the different combination of 50 70 60 80 these combination of uh the different hyper parameter and clearing calculating and I'm passing this to the uh five different model as a parameter and I'm going to try this one and I'm going to test this one and I'm going to print the accuracy this is manually MB I'm not using anything straight away I'm using that so the accuracy of the model one two three four five and getting this one so look at this accuracy and the difference uh with the different parameters so that we have paused over the list right but this is tedious job will be running the N number of C if for example you're saying hey I have only four combination you can do that if it is a 400 combination so in real time scenario nobody will go and do this one that's correct so that's very important thing okay so we have to careful about that this is actually a manual one let's see what's happening so grid search so in the grid search why what I'm doing I'm bringing my grid search here right and I'm passing all this value and I'm highlighting all this uh the algorithm um inside that so I'm passing this parameter uh to my grid search and okay and if this is my model right already I fit I created a classifier model here and passing this grid parameter this was decided here so here I have uh a combination okay I'm going with this so for different algorithms I'm using so after running this one I used to get a eight subsets one two three four okay so four algorithm eight substance so each and every giving a different combination when I look at this one I could understand hey what is this simply it's like giving a one two three four five and one two three four five and we have an algorithm and specify whatever I specified you can understand that right virtual is the best no worries we have a simple function saying that how the views like regression DOT score the same way best underscore params underscore so that will give the the algorithm Auto uh and uh n underscore neighbor 6 is the best one this is getting output okay which one is that um yeah this is the one okay so this is the one actually the grid search is giving the example I mean like giving that the best one so after I'm passing this one so this is all done by my simple grid search option so grid search is simply you can assume uh the hyper parameter space what I given to you guys right so in the grid search in the space the given hyper parameter space the things will go on finding the availability of the combination of uh different data and what are the algorithm we are passed So based on that it is giving and this is also providing me this is the best one okay fine this is the best one how I how we are saying like I'm not relying on even the machine I even I don't like manually putting that all the inputs like in one by one and final I can find out the answer so that will take time so that I'm coming to the machine computer system so I'm doing my mathematics and everything but still I'm not I'm convinced with my result this is telling the best one is this combination which combination Auto n number is equal to six one then let's see that so the result of the mean Square One that's fine which one is the best score from the above list it is then simple already we found the best parameter as this one so compare the 32 combination of this one and you will get the answer 7.73944 749 is the highest value so which one is that from the top one two three four five six this is the one we are looking the system says hey this is the best one for me then I'm finding which one is that again 6 from the top numbering one two three four five six look at carefully all these combinations the other data really is less than this one then I'm finding this selector score which one is actually is giving you 70 that's fine but which one combination is the best combination my algorithms given that values like Auto knife or six is the best one when you look at that account it's saying like 0.73 is the best one so this combination gives me a best result it's a simple one otherwise I'll be really confused so that when I deploy this model I can directly go and apply this one and I'll get my score and model I further things we have a lot of other things we have right I can go ahead on that okay so this is how we used to do that so the same way actually I've done for something for random for us so since we have a nine more minutes so let's catch up the uh questions instead of hanging over here let me see the things right okay who is given that how does one decide a number of components in PC of course that's what I'm telling the when we are building the machine learning that's AIML hold not just like a building not just like because actually we have worked on that even I build like a lot of uh applications right it could be a web application mobile application whatever it is that's real experience and I moved into the analytical and jumping over here that's fine but comparing with the normal application and uh the ml1 we cannot like do do just like that because actually we are dealing with the data that's very important so when you're dealing with the data we have to be careful and we have to do with multiple random I mean multiple way of iterations that's very important that's the reason actually when you talk about AML we have to be very clear like the data needs to be cleaned properly right there are a lot of steps before that that's the reason I said if you ask me like what is the roadmap for the machine learning means it will take for me at least a 16 to 18 months based on the data volume of the data while the data is like as long as you are giving the N number of data to me I'm happy but the data should be a reasonable data otherwise I'll be collapsed all right I cannot find out the one but slice dicing that and will find finding that problem statement I need to get that only the military data and I need what exactly the problem statement fine tune that and get the data from the sources and you do all the exercises and then applying all the stuff so the PCA is a best option but when I'm using a PCA we have to experimental point of viewer to do that so okay so probably like you can find my article uh again I wrote in a year back in uh so how to handle that situation um there are combination of things like that was a nice article I very much interested like I wrote that one so probably you can check that one so that is my answer for you so will slides will be provided let me think about and share with you okay I'll I'll try to do that it's not a like kind of like a what is that a very secret one I'll try to share this one and is the percentage of partition of the data falls into hyper parameter for testing it doesn't matter that's what I said right so so when we are doing with this probably you know right before you split uh the one 70 60 or sorry 70 and 30 75 25 after that you are fitting the data and then you are applying your hyper parameter it doesn't matter because actually when you look at this one let me close this one and go back here yeah see here so here actually you're spitting all your that sorry yeah so you are you splitted your all the data right 34.3 so 70 and 40 for 70 and 30 combination so it doesn't matter actually where like the um for trial and purpose or like when you're learning machine learning that's fine but in real time scenario the most other thing will go to uh actually to like once the model that uh in uh production that's very important okay right all right what is the meaning of optimization optimizing the model optimizing model that's what actually so we'll try to get some value right so if you want to improve the uh your uh your score so each and every uh assessment or whatever the exam whatever you say so you are you are studying well and you are writing exam and you are not getting the score what is the problem you need to find out that so we need to find out that the painful areas and lagging areas so where we need to jump in and we need to figure out that hey these are the like uh the Gap and these are the very important thing we need to look at that one so that way we are finding uh the gaps and improvisation for for example like so training that one like myself for example I want to read some specific content so I need to read myself understanding better way so I'm optimizing myself to perform exam same way when you are doing a machine learning there are a lot of things for optimization in different stages of machine like that's very important not only for this completely am stage so we have to take care of uh right from the selecting The Source data there also you need to concentrate on optimization so you cannot just like that all the data into the model no that's it never end up process after five years only you will get the model right so that's very important optimizations like making you're improvising your performance sometime this is a debatable thing basically it's a debatable thing actually when today I'm giving something it's performing well for example uh I'm running the restaurant so uh some some set of things actually dine in and take away so some sort of things I can do it in dining only something like we cannot do it in decades in earlier days you cannot get the coffee as a takeaway that's very challenging one because we can get all your items like Indian items or anything like bread what are things you can get that as a technique that's fine but now you can get anything I said take away almost 99 percentage I believe right and specifically they customize some pattern to the only for takeaway combo kind of stuff so how they are doing they are finding that the opportunity business opportunities for the how they are finding that based on optimization techniques so wherever possible thing is available so they are improving that business and they are filling all the gaps and makes this strengthen one streamlined process that is nothing but optimization that's very important okay and then how does one design the number of yeah I think we are answering this one and uh again how many play generic algorithm hyper parameter tuning yeah yeah this is simple yeah actually I've shown the slides my slides right Shiva so it's a very simple thing so all the algorithms it has hyper parameter if you have the background of programming language I would say it's a simple syntax and parameter it's a not an optional parameter sometimes it would be option parameter or sometime it's where mandatory parameter so what you used to do like for example uh concatenation string concatenations uh and the string length RS like uh what is like the in some sometime actually finding the extract in the month from the date something like that are extracting the Year from the date so we have a lot of methods in it right I mean a lot of um options I mean like the parameters right parameters like functions so similar to that that's all but we try to understand because we are we are playing with the data so we are applying that the reasonable value for a hyper parameter you will get exactly what output should come based on the volume of the data that's very important so that hyper parameter tuning is like so it's already existing in that the the the subsetter in the uh simple function right and the passing value is in your hands and you have to consult with the smes and pass the values that's a very important thing okay good and even nowadays actually there are hyper parameter um and finding the like a value exactly based on the data we have a lot of tools that we have available so that tool will extract and give that the detail about hey these are the the hyper parameter value we are missing you have to pass this but we cannot relay on time if you are strong enough with the base then you can justify the answer but you can simplify you can you can you can save your time that's very important thing we have but hyper parameter is already there you have to pass the value that's it as I mentioned date function sing function numerator functions the basic function what we have in our in programming language same way here but here it's very very sensitive that's very very sensitive what is the best range of parameter to begin with uh that's depends okay I selected question so what is the best range for the parameters to begin like it so it depends okay that is the one guidelines is that because actually now we are growing with the um all these ML and a cross so that really will give the uh the clear picture about that okay right by using we have got this optimization techniques that's what I'm telling actually so we cannot decide this optimization is good or bad so we based on that uh the data nature of the data and the data has been selected for the uh ml perspective ml tabular hope you guys are know like a tap um uh what is that uh uh what is that one minute let me pick that word um structured data unstructured data and semi structure data if you're talking about all these kind of uh algorithm if it is going to be like a structured data so when you're doing the structured data what are the data but passing the data with the combination of different algorithm will get the output then we can decide hey this is the one thing we can go ahead and so that way we can do said oh there are big questions I've been traveling since Monday unfortunately I lost the session okay that's fine okay so we can get the model parameter paper parameter of the decision tree so this is say that's what I said like when we are building the dysentery there is the options we have that's clearly you can understand that's there are few parameters that there and all the it's going to be Guinea model or other models something like that model parameter if you have education internal you can see that foreign so pretty much I covered okay somebody's searching job I need to be a perfect and rank of our audition everything Mohan actually it's everything so that's what I'm telling you'll try to Deep dive and digest the the concepts and there are a lot of online videos and YouTubes everything we have so just explore that and definitely uh you can get that and same time try to understand the industrial standards so that that will help you a lot okay so deep dive there is a lot of in in turn things that they are giving so play around with that so dedicate yourself for at least a six bands are I would say minimum six months dedication on this for freshers and even people are experienced people they need us at least six months to digest that if there is no programming background that's really challenging but they have to digest that one and come back to this one so it will be really easy otherwise or fake they were they're struggling something like they don't finish that one uh could you please tell me how to decide the range of the hyper parameter to the beginning that is that's what I said right so here we have given that some data I use like for a combination probably you can put a nine combination and even now we have like some other way around like uh um some automated way to build the hyper parameter set so you can have it uh a list of uh for example for parameter let's say sorry I'm taking too much of time that's very useful I believe yeah yeah it's fine yeah okay so when you guys are playing around uh which one that yeah see you later they stick one so what I will do I will I will keep my Excel or see some simple uh simple XML file or simple CSV file the combination of all three in different different scenarios difference different like position and same range might be different so let's say like the hundreds of combination I'm putting in my Excel sheet what I will do I will automatically out is a kind of automation so keep that as a feed for this parameter upgrade and pass through this specific um uh what is that yeah specific uh the grid CV then my complete automation will be good I know I know I don't want to put all the things as a texting say here I'm putting four item and again writing this one the simple way what I will do I'll have Excel sheet I'll keep on adding my permutation information of all the stuff right as a crazy guy so I put that I will get the output that's it simple okay that way you can Implement that it's not a big deal okay um so what is the meaning of the reducing the noise of the model um yeah that's good question basically the noise there might be you know the multi-linear regression model we used to say like Y is equal to uh beta or not plus beta 1 x 1 plus beta 2 x 2 and so on that beta is nothing but beta 1 beta 2 is nothing but a noise you can ask me what is it you're selling the equation yes understood but when you are doing a regularization techniques that's what I said initially there have a regulation techniques to reduce the noise of the data cleaning the data and finding the data wrangling the data had selected different it's a different but when you're doing this kind of Regulation that's what I said like we have a lot of stages we have a data collection data Eda process data wrangling everything Etc and future extraction future selection blah blah blah blah whatever things and to come into the model I'm putting a lot of model relevant to the problem statement and bringing that after that I'm doing hyper parameter tuning after that I'm doing a regularization techniques so why this is completely filtering that rises from my data sometime actually not only that the empty column or duplicate values will not answer all our questions there might be a data too that's all is a laser method right laser technique so that will that will what is that it will squeeze the the data as much as possible to reduce the noise I think you've got my answer right basically provide me the presentation of Jupiter notebook okay I'll try to do this okay it's not a big secret I will share with you guys how preparator regulations later that's what I said actually if you look at that of the equation what I showed you right the sigma of Y your y y i minus x c plus land of w there actually regulations then that Lambda represents nothing but regulation regulation Lambda is nothing but C is equal to 1 by Lambda if you understood that logic okay so I don't know like we have been Crossing six minutes no issues so let me quickly bring because otherwise uh will be which one that yeah so this is your hyper parameter and this specific thing actually I can call it as a regulation techniques I can use this there what is that what is your question questions that is the model needs yeah yeah regulation regulation related so that is the regulation related that's Lambda right when you talk about like regulation techniques this specific equation come over there then the Lambda will be playing a major role and the weightage will be going a very important thing so that is a combination of that one okay uh what is the meaning of it is okay right that's fine and but that if the data is wrangled what noise you are talking about it data is wrangled is just clearing the data cleaning the data on top of that noise see if I'm speaking right I'm speaking a right words and right sense but there might be like some internet issues or a signal problem my side are your side that may be signal issues that really there is some noise in in in in communicationally in electrons and communication way I'm communicating that one it's a similar that one so probably you can see probably for speaking with someone everything is fine your mobile is good that mobile is good but when speaking you're not able to hear some some background noise will come out right so that's not that that is nothing but after doing all the data wrangling and all still there is a scaling up scaling down we have a like a scale up scaled on maximizing your data something like that so we need to handle that one so that is a kind of a noise again regulation techniques that will reduce the noise uh better way okay so I think I covered pretty much all the questions so you guys can reach out me through uh my LinkedIn uh way and uh even I have my regularly writing all my articles across um and their analytics with Dahlia and data Science Central and even I have my own blog shanta's AI View and I have my YouTube channels so probably you can reach out that City teacher is a little question only have I need to build that but you can you can join you can connect me through any social media okay all uh your questions and collaboration Works will always be accepted noises outlayer outlayer is different outlayer is totally different outlayer is totally different so we can you can see that outliers necklace noises is like you're not you can you cannot see your normal eyes simple way outliers when you put that the uh blah what is that box plot you can find out that you put that some Holograms uh view of Eda process so that you can find out a this is not fitting to my my belt so remove the data that is outlay is different noise is something hiding from your eyes that is very important okay so from that you can you can find out that that is different noise is different from daily from outline it's outlier we will do all the things in our still that's what I say today the model is working fine and it's after three four days the model will go on down because the data the data is working fine then that is nothing but data Drift We have a data Drift We have a model to so many things we have so it's a long process I think already uh more than 10 minutes so I hope uh I covered all the questions and answered your questions I believe uh yes sir yeah so thanks a lot on behalf of analytics media I would like to thank you for your time and for delivering such a wonderful session and I'm sure our audience find it insightful and hopeful uh we can conduct more session with you in the future thank you definitely definitely okay guys I think hope you understand that tuning so let's go back and perform and Improvement your model let's catch up with next session let me get the time and we align with you guys thank you thanks everyone and thanks for the opportunity thank you yeah bye