Credit Card Fraud Detection using Machine Learning from Kaggle

[@snehalbogar6326]

Thanks!

[@karthikdeepan1998]

As this is imbalance dataset, accuracy metric doesn't work. We have to use confusion matrix, recall as metrics.

[@muhammadwajahatali2286]

Great bro❤

[@swapniljena8684]

Hey I have a suggestion for improvement of recall. As this is a class imbalance data to detect the fraud transaction so accuracy is not the metric that should be counted on. The best recall score you have got is 0.27. I got a recall score for class "1" i.e. fraud 0.74. Using the random forest ensemble method, further when I used decision stumps of random forest and trained it using Adaboost I got even better recall of 0.76. Thank you for the video, got to learn a lot.

[@dineshramachandran1961]

@Puneet Rajput , Im impressed with ur knowledge. I am aspiring DS. Can i get ur email id pls

[@akshats5996]

So true!

[@manthanrathod1046]

never heard about this method of using decision stump and further boosting it. Although I used plain smote enn and got 82% f1 score. Can you redirect me to video or site? I need to study about this decision stump and ada boost technique. Hope you will reply. Thanks :)

[@divyanshumishra5993]

@swapnil jena Can you please help me to create the front end(GUI) for this project as web application??

[@snehalbogar6326]

Thanks for explaining things in really simple language, yet covering all the complex theory.

[@ayush51379]

Thanks a lot for sharing this very useful video :-) I would like to add that we tested the model here only on the training data set itself, while this is an important step, the next step is to test the model on a testing data set. Probably, splitting the data set into 80% training data and 20% testing data will help. Also, one should perform error analysis and hyper-parameter tuning to get best results, taking care to keep the model generalized enough to accommodate for the entire data set that also consists of the data set given to us. Have a nice day! :-)

[@divyanshumishra5993]

Can you please help me to create the front end(GUI) for this project as web application

[@sankarsai5054]

finally, find my...final year project on youtube

[@adityatiwari2488]

Are u going to create a web application for the project?

[@sankarsai5054]

@@adityatiwari2488 no

[@adityatiwari2488]

@@sankarsai5054 bro i am facing name error

[@adityatiwari2488]

In jupyter notebook NameError :- name 'Fraud' is not defined Can please help me what should I do?

[@sankarsai5054]

@@adityatiwari2488 in just..buy my project 🤗

[@otmaneelaloi7926]

Accuracy sccore isn't a good metric for this task since the data is very imbalnced, you can check the precision and recall you've gotten for isolation forest and local outlier factor. A good approch to compare theses algorithms on this specific task would be to use auc_score. Thanks for your vedios.

[@magicmushroom9670]

exactly.

[@kimchi6284]

hey , i have a project in this topic and i don't have any idea about it could you help please

[@nashgaming2761]

is there a dataset that shows the variables used, without confidentiality issues

[@Han-ve8uh]

11:31 says "root node will be selected in such a way that the outlier will be splitted". This doesn't sound like ranodm selection of split point, which contradicts what the text in notbook is saying? "by randomly selecting a feature and then randomly selecting a split value"

[@lanashin2631]

Wow you explain it so well. i learned a lot from you, thank you!

[@baidash3104]

Fraud data:492 Normal data: 284315 If we use a dumb model and set it to predict all transactions are normal then we will be getting an accuracy of 99.83% which i believe both models are doing internally based on the Precison and recall of +ve class label..

[@piyapiyagill]

which dumb model?

[@baidash3104]

@@piyapiyagill Wont need a model. Given any query we just say its not fraud.

[@jimmywhite5458]

Great explanation- super interesting thank you

[@jammulanarendar9910]

Hi Krish , I have a small doubt ,if I observer the Classification report of Isolation forest ,Local outlier factor and SVM ,The Recall is very less for 1(i.e fraud), can we say the model is accurate only by just looking at the Accuracy (as per my knowledge accuracy is not a good metric for class imbalanced data because let say sample of 100 records having 90:10 Non Fraud : Fraud records even if model blindly say everything is Non Fraud then the accuracy would be 90%.) That's why Kaggle people Recommended to use the AUPR Curve as performance metric for performance check. Thanks in advance.

[@pawansapkota3970]

Yes Area under the precision curve is the best metrics for the given datasets. I am also doing the project in the same dataset so balancing the datasets using SMOTE and then classification with the Random Forest yield a better result.

[@divyanshumishra5993]

Can you please help me to create the front end(GUI) for this project as web application?

[@thisaintarf]

why you dont use sensitivity and f1 score to evaluate model perform than accuracy instead

[@Amina-xu8uj]

Thank you, nice explanation

[@mbmk92]

Thank you for this approach regarding the imbalanced dataset issue for fraud detection. However what you opinion in using oversampling with Synthetic Minority Over-sampling Technique (SMOTE) together with edited nearest-neighbours (ENN) (ENN) instead of Anomaly detection approach?

[@fekiyounes5181]

I think that if you were able to generate data you are able to classify them... :D UnderSampling is the other side of the coin, you are creating a boundary by eliminating the most important and useful data that you could probably misclassify if not deleted... In my opinion the best options are sampling and creating voting classification or autoencoders and loss thresholding....

[@meetmeraj2000]

why didnt you use SMOTE for upsampling the data?

[@mishralucky]

How authorised push payment fraud prevention and detection can be implemented using ML/DL.. please share a video if possible Krish

[@jigneshkhandare321]

I want to make project on this topic using different algorithm can you help

[@ANJALIVERMA-d4c]

I wanted to know why everyone is using only this dataset as on the kaggle websites lot of datasets are available for the same name "credit card fraud detection" but wit different data. Does anyone knows, can I make my project on different one also ?

[@rishabhjain2559]

how can we say isolation forest and LOF are performing well? Recall values for class 1 are very low in both cases. model is not able to detect even 50% outliers, accuracy won't be a good metric, because of the imbalanced data

[@motivational_19171]

IN[27] i am getting 'not all arguments converted during string formatting' this error

[@shikharsaxena9989]

thanks for clearing my concept

[@swaniketchowdhury]

This may sound crazy but how can we check if a transaction is fraud or not by using some other variables like card number or something like that?

[@doomsday7699]

If you can convert such variables into numerical values that have meaning, then you can try to make a correlation table and see if the output is dependent on such variables. That is one way of doing it. Another way is, if you feel that a variable is not useful, you could directly drop it and then see the accuracy. Or there might be another statistical way that I do not know about.

[@MrChudhi]

Just a question which metric is the best metric to detect the credit card fraud. I believe it should not be accuracy. So it should be recall. Am I correct.

[@omkarr8282]

Hats off man!! I this is one of the best videos for Imbalanced datasets, i was tired of reading about SMOTEs every time i wanted to look for ways to deal with taget variable imbalance! Thanks for sharing :) I was wondering, for training , did you stratify the data ? does it really matter if we stratify the data or not because it is already imbalanced?

[@sgracem2863]

SMOTE actually seems a lot simpler than this though?

[@sneha2502]

Last part of the code is not working sir.. 😬.. it is showing that random_state is unexpected keyword argument... What i have to make changes.. please let me knw sir.. please

[@ajoychatterjee3105]

Accuracy is not a factor here to concern. But precession and recall are 100% accurate for non-fraud . But only 0.26 and 0.27 for fraud transaction. Is it a good number to consider ? Just curious to know what real time business accepts.

[@meghasmita152]

This is on 0.1 percent of data, results could vary with other data samples, how to effectively sample in such situations?

[@tanzinahossain8179]

Hello, can someone explain about v1 to v28? And how many features are taken And also the heat map...

[@ashwinimagar4822]

Thanks for the explanation. Why recall in your result looks really low?

[@krishnaik06]

It is an imbalanced dataset.

[@ashwinimagar4822]

@@krishnaik06 Yes I got that but generally cost associated with false negatives in fraud detection is very high. Marking fraudulent transactions as non-fraudulent is expensive and hence recall for fraud class should be high. What is your cost function here?

[@prashantsolanki007]

@@ashwinimagar4822 Yeah his whole metrics is messed up. That model is no better than a dumb model giving all negative.

[@quaziharisahmed6172]

@@krishnaik06 sir whether it is class imbalance or balance class what is the use of model if he is not able to predict properly ?? in this case random forest is performing much better than isolation forest

[@quaziharisahmed6172]

@@prashantsolanki007 true

[@naveenmami7438]

Thanks for the inputs ! will it exclude the anomalies detected before building the model with isolation forest and calculates the accuracy? please elaborate sir

[@divyanshumishra5993]

Can you please help me to create the front end(GUI) for this project as web application

[@hariharangr4045]

Thanks for explanation.Can you explain how to remove the detected outliers in the dataset?

[@eswarsaipallapolu7484]

from pylab import rcParams rcParams['figure.figsize'] = 14, 8 RANDOM_SEED = 42 LABELS = ["Normal", "Fraud"] For what purpose is this block of code used? Can any one explain it line by line

[@divyanshumishra5993]

Can anyone tell me, how to create the front-end of this project?

[@divyanshumishra5993]

Pls... Help, if you can..

[@muditrustagi5775]

sir i have a doubt regarding a code cell that you have provided

[@bharathkulkarni4207]

How will it really reduce the fraud??, even the user himself can report to the bank about his unnotice transaction , what action will be taken next?? Literally am unable to understand how will this project reduce fraud, it will just identify the fraud which can be detected by user himself. anyone please reply.

[@rakeshvanga5514]

would you pls provide me the link for dataset

[@TT-ds7kb]

And what it will be considered as I mean as a software or application

[@gungunalewithsakshi7579]

Could you please explain that if we have any transaction how we will get know that this is fake transaction, if we don't have any feature like class (0, 1)

[@vivekkumar-dn2vb]

How to resolve this error sir plz help In[25] : __init__() got an unexpected keyword argument 'random_state'

[@kislayanupam1027]

Set the 'random_state'= None or remove it.

[@thecryptotradingclub]

Nice Video. Thanks for sharing.!

[@bogdanilie6152]

Why are you using SVM? Isn't this an algortihm for Supervised Learning? Given that we have unlabeled data, I assume we should not use it. I'd greatly appreciate your explanation. Many thanks

[@kushagrasharma4075]

It is labeled

[@shreyasbhosale5651]

Where do i get dataset for the code?

[@sakshimishra6450]

on kaggle

[@satyabansahoo6075]

Sir, can I fork your projects and work on that?

[@lizmathew1481]

Sir,where can i find the source code for this project?

[@chiragrana5039]

I still feel that if a bank has an fraud of very high amount it won't include that in the dataset as it might create a hugh problem for the bank, so ho w much to trust the data is difficult to predict.

[@tyitb156shubhampakale5]

Where can I get creditcard.csv file

[@aniketgaikwad1157]

Do we have to fill the null values before applying the ML algos in this video???

[@doomsday7699]

Yup. Else will always mess the entire algorithm up. Might even throw exceptions depending on the library you use.

[@rh334]

Code not working. Declare input variable first

[@jen_1105]

sir where can we download this note file ?

[@louerleseigneur4532]

Thanks

[@TT-ds7kb]

Sir where we will deploye the project

[@nashgaming2761]

Can I get a dataset with principle parameters as transaction location and user patter(behaviour) like daily transaction amount etc

[@swapniljena8684]

All the datasets on kaggle are PCA performed on them related to credit card transacion

[@nashgaming2761]

@@swapniljena8684 is there any other site where I can get??

[@swapniljena8684]

@@nashgaming2761 Well you can try to search to get raw information but I don't think it is easy to find. No banks would post out such information without encryption.

[@prakritsinha3094]

can anyone explain why are the metrics for class '1' so low? eg in isolation forest algorithm why are the precision, recall and f1 score 0.26,0.27 and 0.26 respectively? is that the desired result?

[@annapoornayaligar4187]

Sir this was very usefull for my presentation but can you please tell me how you will predict the card is frauded whether by scanning card or using transaction history plz reply it's very important to my presentation

[@datatorture3086]

Dataset has already defined which transactions are fraudulent but sir has used unsupervised learning algorithm to carve out the relevant information pertaining to the dataset better you should explore pycaret module that will simplified all the preprocessing steps

[@bhanuprakashreddyvennapusa7790]

For checking the Transaction, Where we can get that Transaction Details

[@bhanuprakashreddyvennapusa7790]

@Krish Naik Can you reply for my comment and question

[@shahalam-ei4cm]

sir from where we get csv file plz .....tell

[@vinuyesudas5396]

Kaggle.com is the best site for various datasets

[@sgracem2863]

www.kaggle.com/mlg-ulb/creditcardfraud?select=creditcard.csv

[@shubhangisakarkar9532]

Hello Sir, On what basis the dataset is imbalanced.Please elaborate.

[@doomsday7699]

The number of examples per class in the dataset. Number of examples for the fraudulent case are much much lower than non-fraudulent examples. Balanced dataset means equal or close number of examples per class.

[@sandipansarkar9211]

good

[@jaysoni7812]

why our classifier return pred value -1 and 1 instead of 0 and 1 ????????

[@chamodmaduranga3866]

its default return.If you want to change to 1's for fraud and 0's for non-fraud, use map function.

[@jaysoni7812]

@@chamodmaduranga3866 but why, for other problems it give same value as in y anything special in this case?

[@ashwinikattimani9298]

How to train this data using tensorflow lite?

[@ashwinikattimani9298]

And how to add it to the android app?

[@mithunkumar7063]

How to remove those outliers from the data

[@__ALahari]

How we know whether it is fraud or not

[@sujithpawan7246]

How to download ipytnb of your data

[@ramanjeet1111]

companies are still using the manual methods

[@redditinside.8722]

bro i need description of these project

[@magicmushroom9670]

kzbin.info/www/bejne/nKOwkGqLgqmSY6M at this point is giving you that there are 73 errors but those errors are from both fraud and normal transactions. I researched and in my opinion this is not a very good metric for checking accuracy of an anomaly detection system. You didnt explained the Recall which is most important part of this whole video. Apart from that it is more helpful if you explained the ROC and threshhold in this scenario which is what is needed. not just implementation of an algo. BTW I have great respect for your work and love what you are doing.

[@shrenikaadsul1154]

data=pd.read_csv('creditcard.csv') data.head() In this command give a syntax error

[@ruthwikakumari3622]

Can I have a vedio on random tree based and cart based algorithms for credit card fraud detection

[@divyanshumishra5993]

Can you please help me to create the front end(GUI) for this project as web application?

[@akshats5996]

This is a classic imbalanced dataset problem! You lost the plot when you focused on accuracy instead of Recall. Recall is very poor in the outputs you have shown.

[@varshatn2537]

@krish naik

[@sudarshankadge7945]

How to detect fraud transaction with same file without using "Class" column.

[@jeevanchavan143]

Using unsupervised learning,clustering

[@swapniljena8684]

First we need to know how do we classify a transaction as a fraudulent one, then only we will be able to know how to detect them.

[@swapniljena8684]

@@jeevanchavan143 How do we know which cluster belongs to the class "1" i.e. fraud

[@shravanKumar-yc9cj]

after clustering in the know data ,u can use in the the file with unknown class

[@varshatn2537]

Hello sir, we are currently working on this project... We want to get in touch with u for more help and details... Cud u pls provide us with your email .

[@vinuyesudas5396]

How can analyze transaction is fraud or not without class? This dataset already contains class as transaction is fraud or not info ryt?

[@manankalra7978]

@@vinuyesudas5396 I guess, knowing the Class variable has no role in training those models. Class here is just used to calculate the accuracy of our predictions.

[@chamodmaduranga3866]

@@manankalra7978 contamination in isolation forest is used based on a class variable.