I think this question should be tackled in a hierarchical manner. For example, 1. First, check does the data have a label column- if yes (go with supervised learning) else (go with unsupervised learning). 2. Second, check is the problem prediction or classification. 3. Third, check size of data. If the data is too large go with ANN else Machine Learning 4. Fourth, apply basic algorithm such as linear, svm, etc. first 5 Now, apply ensemble techniques.
First we will see the type of problem. supervised/unsupervised? classification/regression?...then we will try out multiple ml models for each case and see which model gives best accuracy
Visualisation of data using seaborn library ,if my data point are linearly separable then I use linear regression and logistic regression if not the I go for ensemble techniques , and SVM
@krish Naik can you please create a playlist required to crack data analyst role. You have given the roadmap. Requesting you to upload a playlist required. Thank you in advance Krish 😊😊
Hi Krish...even i faced the same question in interview and said i will go and apply all the algos and select based on trial and error and will select the best model with good metrics.. but the interviewer said that it is not proper approach to select a model on trial and error ...
we should not say trial and error.. but we can say we have model evaluation techniques like the confusion matrix, classification report to determine which model is performing better in that particular data set. its always better to check with 5, 6 models before coming to a conclusion about the data set. I guess you were correct the only thing was not to mention trial and errr thing :)
That depends on the algorithm..if it's linear regression, it uses least squares to determine the line of fit. It will calculate values that minimize the square of residuals from tht line. If it's logistic regression, it follows max likelihood estimation..if it's decision trees, the algorithm follows entropy or gini index and node impurity to calculate the best split and goes until the depth you define. So each algorithm has its own set of criteria and methods to follow.. these are some of the ones that I know. Correct me if I'm wrong.
Yes, they helps in finding best hyper parameters available in your parameter space. But again it depends upon your parameter space, how you are choosing that.
Hi Krish in one of my interview I faced questions like. 1. What is the loss function in collaborative filtering 2. How we can use regularisation techniques in collaborative filtering Can you please explain
I think this question should be tackled in a hierarchical manner. For example,
1. First, check does the data have a label column- if yes (go with supervised learning) else (go with unsupervised learning).
2. Second, check is the problem prediction or classification.
3. Third, check size of data. If the data is too large go with ANN else Machine Learning
4. Fourth, apply basic algorithm such as linear, svm, etc. first
5 Now, apply ensemble techniques.
How do you decide weather the data is large or not ?
@@veeresh4441 by number of records and features....ultimately size of data
@@reenasheoran893 suppose you have some 10000 rows and 100 columns or features , do u consider this big?
@@veeresh4441 no
i think this is a smart way to answer.
most important and common question in every interview ✌️
First we will see the type of problem. supervised/unsupervised? classification/regression?...then we will try out multiple ml models for each case and see which model gives best accuracy
I prefer your old style of teaching rather than using whiteboard/blackboard. It's a good question though!
Visualisation of data using seaborn library ,if my data point are linearly separable then I use linear regression and logistic regression if not the I go for ensemble techniques , and SVM
Perfect 🤞🏻😎😎
@krish Naik can you please create a playlist required to crack data analyst role. You have given the roadmap. Requesting you to upload a playlist required. Thank you in advance Krish 😊😊
Hi Krish...even i faced the same question in interview and said i will go and apply all the algos and select based on trial and error and will select the best model with good metrics.. but the interviewer said that it is not proper approach to select a model on trial and error ...
we should not say trial and error.. but we can say we have model evaluation techniques like the confusion matrix, classification report to determine which model is performing better in that particular data set. its always better to check with 5, 6 models before coming to a conclusion about the data set. I guess you were correct the only thing was not to mention trial and errr thing :)
I also follow the same approach which you have suggested.
It depends on type of data then we can only deside which algo to be applied..
Sir please make video on retraining approach based on performance metrics
STAR should be used to structure the ideas
Basic ML interview question: Why/When ML model stops training after calling fit method ??
That depends on the algorithm..if it's linear regression, it uses least squares to determine the line of fit. It will calculate values that minimize the square of residuals from tht line. If it's logistic regression, it follows max likelihood estimation..if it's decision trees, the algorithm follows entropy or gini index and node impurity to calculate the best split and goes until the depth you define. So each algorithm has its own set of criteria and methods to follow.. these are some of the ones that I know. Correct me if I'm wrong.
@@pravali4452 you said that is correct but that is not answer to my question
When error value(loss) doesn't change much after each training iteration? Is it right? If not, may I know what's the answer!
Great video krish naik.
How to resolve 'XGBoost error: bad allocation' in multiclass image classification? Please let me know.
Search in google
But applying every algorithm won't it be time consuming?
Auto ML models does that for you. It applies every ML algorithm (a bit consuming, though not much) and gives u the result.
Hi Krish, How to check if the data is linearly separable or not ?
You can use pair plot, in short visualize the data first and then you will get to know whether it is linearly separable or not
Please advise
Can gridsearch or random search fix overfitting or underfitting issue?!
Yes, they helps in finding best hyper parameters available in your parameter space. But again it depends upon your parameter space, how you are choosing that.
If that is the answer, won't the next question be it is very time consuming to do trial and error ?
Model needs to be created in that way
@@krishnaik06 I thought your first approach "after performing EDA, selection of model" was the best one.
Yes but now if see the automl models they verify with all the algorithms
@@krishnaik06 Yaa right. Microsoft Azure does the same thing.
@@shanbhag003 Why should one worry about choosing right algorithm when we have AutoML!!
if i use pycaret and comparing the model , Will it be enough?
Can anyone suggest me a book to refer for Python learning. I am new to this course..
Crash course in Python is a good book to start.
Better to check python documentation... best way to learn python
Good one👌
Hi Krish in one of my interview I faced questions like.
1. What is the loss function in collaborative filtering
2. How we can use regularisation techniques in collaborative filtering
Can you please explain
Your presentation looks like you're imitating applied ai course...srikanth sir class 😅😅....great job still
:) I really adore him
@@krishnaik06 That's wonderful and we adore you ...please do your traditional kindo videos which you used to do at the beginning....thank you 👍
Don't think no one can be compared with Srikanth sir..he is in different league
@@vyshagha3747 There's no comparison which im making....we are already standing on the shoulders of Gaints. Its my observation which I expressed.