What is Data Leakage In Machine Learning?

[@JS-gg4px]

You are presenting like you really want everyone to know the significance of DL. I like this attitude! Thank you for sharing.

[@nitinpatil6611]

Jjj

[@erfanm5276]

I love your passion in teaching man, well done to you. it was a great lesson, concise and precise

[@mthandenimathabelacap5466]

Very intuitive and good that you bring knlowledge about the implications for production stage of the model. Great.

[@aminderkhungura7929]

Thanks!

[@mahdikhalilnejad]

Hello Thanks for your excellent explanation. I watched several sites and videos until finally your rescue video helped me understand this ... don't get tired, you put a lot of energy into it.

[@sdc8591]

Well that was very quick. Just sometime back in Live Q&A someone requested for this information and it is here within just few hours !! 👍

[@ismailyt6627]

That was what I was looking for, an excellent explanation with simplicity and schemas, thanks a lot

[@kianestrera-hr5vt]

Wow, and that's it most of my questions and needed to clarify with data leakage is answered within 10min, Thank you so much, wish your channel grow much further

[@saurabhmukherjee850]

M a big fan of yours dear Krish... Your contents and way of presentation is simply flawless... Keep on doing the good work... God bless you

[@FrancisKyalo-hb5pu]

I really enjoyed how well you explained the concept of data leakage, Thanks Krish Naik

[@Neojs565]

Kaggle brought me to this video....wow man...really eye-opening..!

[@GurtDovletov]

Thank you so much such a good video, it easy to understand. Explained with easy examples

[@shivanikhandelwal8324]

Thank you. This video solved all my doubts...

[@4urandy]

This is so legit ... Absolutely production knowledge

[@TuanKhai298]

best teacher ever, thank youuu so much

[@muralikrishna7617]

Hello sir just now i can accross one of your videos i do nd that what your are giving is good

[@manishagyaan]

Great explanation Krish... You are doing great work... Stay blessed always

[@Jobic-10]

Take these medals 🏅🥇🏅 🎉 Subscribed ❤

[@victorahaji9717]

You a wonderful instructor.

[@faustopf-.]

Super informative and very well explained!

[@nitayg1326]

3 valid observations - never thought about it!

[@vinaytomar5827]

Thank you sir for these important instructions 🙌👌

[@alexanderdushenin7035]

2 days ago I meet with this problem on competition. Thanks for explanation)

[@claudineievangelistanascim2562]

Thanks 🙏 for sharing. Very Important Point ! Very much helpful !

[@chaitanyakrishna5873]

Excellent Krish

[@laythherzallah3493]

Thank you ..very fruitfull information

[@vaibhavkhobragade9773]

Thank you! This is a very important topic you've covered.

[@sanjaytallolli]

Very much helpful...

[@lotmoretolearn-dataanalyti9312]

Very nice explanation krish.. Easy to understand. I always follow ur videos. Please do more such videos on data science and data engineering as well.

[@khadidjabenchaira6004]

Thank you so much from Algeria ...

[@ameermohamedr4982]

Hello Krish, please make one video regarding data leakage with simple dataset for deep understanding for us

[@marekdudzik6779]

Best explanation!!! THX

[@saurabhtripathi62]

very imp point thanks a lot

[@sudehashrafi7040]

thank you so much, you explained it very well😍

[@sandeeppanchal8615]

Another scenario- before splitting the data, suppose we 'fit' vectorizer like BoW or TFIDF, etc, on the whole data set and after splitting, we will 'transform' both train and test set. Since the entire data has already been fitted to vectorizer, after split, transforming test set means data leakage. Only after splitting, 'fit_transform' should be done on train set and 'trasform' on test set (also CV set if created).

[@xianglongchen3088]

Thank you for sharing.

[@kevinmatip9416]

The future lies with projects like the Utopia ecosystem because there is a lot of demand for privacy and security of personal data online.

[@jijie133]

Great video!

[@shubhamurolagin5466]

Sir, lovely explaination

[@Abhi-qi6wm]

Good video. An alternate title could be Train-Test contamination.

[@nakshatranahar]

Great Bro

[@jackyfree5756]

Thanks for sharing sir!

[@robertpatterson6445]

Thank you, very helpful!

[@bhavyaparikh6933]

eagerly waiting for next video!!!!

[@Mad20240]

my discovery of the moment

[@MathPhysicsFunwithGus]

Great video!!

[@Abhishek-jy4ul]

This was very informative thank you sir!!

[@bhawnajyoti2000]

Nicely explained

[@Raja-tt4ll]

Very nice video

[@ArunKoundinyaParasa]

super explanation

[@nicolasmarcos5996]

Wonderful explanation. Congratulations! I'll subscribe at the channel.

[@rushikeshbulbule8120]

Excellent👍👏😆

[@MrMandarpriya]

thanks you very much Sir for the explanation. So then what could be the receipe. As in Finance/Economics we have the full the dataset, so then is it better to divide the data based on timeseries(date) and then do the ususal or is there other approach.

[@thulasikumar770]

Well cleared my doubt

[@afzalhasan6494]

u r legend ❤

[@heenagirdher6443]

@krish pls explain the concept of data leakage in deep learning

[@madhureddy5328]

I am a new learner in data science course. In time series analysis when we split data. Do the data is splitted (test data) accordingly required. Because my doubt is like when we split data randomly there is a chance of getting a old data in test dataset and new data in training data. So that may leads to decrease in accuracy right. Is there any way to split data accordingly that can cover equally through out the data set at every corner equally splitted or like covering latest data as test data. Not like covering only latest data or old data. As you are saying we train latest data as testing. Can we split in this way. If possible can you make a video related to splitting data.

[@dhonidamaka5013]

One. Suggestion is test should be chosen wiiith respect to time. Since time series is not random and rather temporal. So split should also happen temporally . For example is u have built a model for data pertaining to Jan-oct. Test should be Nov and dec

[@AshrafVideos]

Hi.What are the differences between data leakage and overfitting of the model and the baising of the training data

[@BiancaAguglia]

1. Data leakage happens when your train set and your test set share a common value for a specific feature. Krish explained it perfectly with the mean example. Let's say you try to create a model that predicts the score a senior high-school student will get on his SAT. Your dataset is the senior-year grades of 1,000 students from different schools. For some of those students, you have missing grades, so you replace the missing values with the average grade of all 1,000 students. After that you separate the dataset into training and test. See how that will be a problem? All the students in the train set and the test set have the same value for their missing grades. Depending on how many missing values you had (in other words, how often you had to use the average to replace missing values), and how important the feature that had missing values was in the training of your model, data leakage can make your model perform very poorly when it is fed new data. 2. Bias is when your model isn't able to pick up all the true patterns. It usually happens because the model is too simple. As a real world example, let's say you're trying to predict if a candidate is good for a ML engineer position. A very simple model would just use the age as a feature and predict that a person over 50 is not good candidates. That's high bias. 😊An unbiased model would train on more features. 3. Overfitting is when your model picks up the true patterns in the data but also all the noise and all the outliers. It usually happens because your model is too complex. Using the example above, a model would train on all the features related to a candidate. I'll be silly and exaggerate here just to make a point, but let's say that, in your train set, you have data like what books the candidates read, what movies they like, whether or not they drink coffee. Most of the data from those features is noise. 😊

[@kunadha]

@@BiancaAguglia Thank you for your time. you explained so well.

[@yashodhansatellite1]

Salute

[@shalakamhetre8093]

Good topic

[@haneulkim4902]

Thanks for great video. So when I have great performance in all train, valid, and even test dataset it data leakage is the only problem? Are there some other problem that lead to such issue? Currently my model is doing too well on all of train, valid, and test dataset however I know from experience that performance on testset shouldn't be this good. I've check all the data preprocessing steps and I did split data in temporal fashion, scaled or filled missing value for each train, valid, and testset independently. I'm still having too high performance on testset. What other issues can there be?

[@SonalAtale]

Hi Krish, I had a doubt for time series. I am having a dataframe with 2000 datapoints (spectral data-Xs) out of which I am having Y values for only 30 datapoints which i am using for model training and testing along with 3 4 more similar subsets. Now if i try predicting the remaining 2970 values, will the model be biased as it is a time series data? Or i can consider it as good as an unbiased test set?

[@Actanonverba01]

Thank you

[@user-cw7yi1ew1z]

So there is no way to prevent data leakage on cross validation then?

[@adeboolusegun1857]

What's the difference between data leakage and data Lake. Also I have not seen any video where you make use of feature selection in practice. I watch video where you mentioned all the types of feature selection but no practical project. I would like to see how you perform feature selection in a dataset with about 150 features. I beg for this, please do a video on it sir.

[@sabyasachidas347]

Thank you :)

[@rusiano-kq9ws]

I think that in the test you should still replace with the mean with the train, not the test. Isn't that so? The same applies with scaling, if you applied for example some standard scaling to the train, in the test you will apply the same scaling using the mean and the std of the train, no?

[@nijagunadarshan2529]

are you saying we should not preprocess the test data?!

[@rusiano-kq9ws]

@@nijagunadarshan2529 no. what i am saying is that if in the train you applied some transformation that involved/took into account the values of other samples (say normalization, or null imputation with mean, for example), then when you get your hands on the test you must apply those transformations using values from the train. This means, for example, filling nulls with train mean or normalizing using train mean and std.

[@vamsikrishnagannamaneni912]

yea true

[@ankit689]

Shouldn't you replace with train mean rather than test mean for test missing?

[@kunalpandya8468]

So, outliers handling, feature selection everything should be done after split...????

[@ThePentanol]

Thanks

[@SR-ng5kw]

I have a doubt here , what if train mean and test mean are totally different. Wont the model learn from train mean and then fail on test data. I beleive the universal mean is a better value for missing value imputation as it represents the global distribution. If your test data is totally different from train data anyhow your model will fail in production.

[@gamer-shub]

sir please make practical videos on recurrent neural network

[@avinashreddy1911]

Nice observations. But what about Encoding for categories?. If train data categories (A,B,C) don't contain test data categories(L,M), we usually combine the train and test data for getting all the categories. What should we do it this is kind of data leakage?

[@ashishsaman]

Stratified sampling? Also, there is no issue if we perform One hot encoding on the entire data set.

[@sulagnanandi2024]

Hi , can anyone tell me does feature encoding like OneHot encoding and ordinal encoding on the whole data set lead to data leakage?

[@manikandannatarajan9665]

Krish, Thanks for your video. I have one doubt. During model selection and finalization we split the data and find the model. But for deployment, we combine all the data to create the newly trained model for deployment. Will this data leakage happens there?

[@AshishBaidyanathan]

I guess K fold cross validation should take care of that

[@HappyBelly10]

Really informative. Hilarious that you purposely used data leakage to improve Kaggle results. Shows we shouldn't necessarily use winner's code.

[@harinishre664]

Should we do EDA only after train set after the split?

[@antifragile01]

Dont you think that global values are better representatives when it comes to imputation? Hence the question that how is a initial split justified?

[@apoorvshrivastava3544]

Sir I have big doubt cross validation can also lead to data leakage because we already seen that data in first iteration and again in second iteration K we given that data in test set

[@gauthamsreekumar5358]

In cross validation, we train K (in K-fold) different models each with train-valid split. That is , In each iteration, we train new model, predict with validation-set and get the accuracy. That’s why we get K accuracies( length of the output array is K) as output. Then we take average of all those values.

[@mehrdadkazemi3969]

thnks

[@bintangmuhammad7082]

So Data Leakage on missing value will happens if i'm doing imputations to replace the null values right? what if i just remove the rows with missing value(if it's only a a small percentage), does Data Leakage will not be happening(because i dont do replacements) or it will still happen?

[@pritampatra6077]

then u may lose some important information if the missing values are more in number.

[@shubhammishra-ht2oo]

Hello sir , what is data mismatch in machine learning?

[@hardikpatel-go2ko]

I am new learner. I was supposed to do missing value handling first then process for train test split which was wrong method.. so pls can you tell us exact steps from 1 to last...

[@favourchiemelaonoh9869]

split first, then handle missing values.

[@AdityaBhagavatula-i1w]

god tier stuff

[@welcometooaudioland7877]

This guy deserves an award. If you only knew how difficult it was to find this topic and someone who nails the issue on the head.

[@336_saranyamaity8]

so thats how some people getting 0 MAE -_- in Kaggle , damn

[@raghavendra2096]

There is no better way to explain data leakage!!!!!!!!!!!!!!!!!!

[@jawad237fx]

so I'm suppose to do things the wrong way in Kaggle amm interesting

[@devkumaracharyaiitbombay5341]

kon kon dl ke playlist ke cnn wale implementation se aaya

[@muralikrishna7617]

Sir can i expcet avideo or a simple message to me from you about free website where we can learn hacking i would be very happy if you reply tot this comment thank you sir

[@lilianasanchez602]

Please translate from Indian to Spanish