Hi, Thank you for such a nice presentation on ML Explainability. I am learning ML, and so far I understood how scikit-learn library and other different Python libraries have made life easier in dealing with any kind of data. Now, I am learning ML Explainability, thinking to build a project with the data available in the public. Can you suggest me which dataset/s would be more suitable for me to start with as a beginner?
Hello and thanks for following my channel. Yes biased data sets are very common in e-commerce data sets. So search for e-Commerce data sets in Kaggle or the UCI Machine Learning data sets are great as well.
Hello and thanks for your question. So lime is a way to explain the weightages or coefficients of non linear models as a linear model would do. That’s why the feature contributions are estimates only which makes them rounded off in certain situations. Hope this helps and stay tuned
@@AIwithSohini so if the predicted probability of a certain class in a multiclass classification problem (with LIME) is 0.03 and the sum of the contributions is more than 0.10 (as in my example), the reason is because the digits are rounded?
Hello there. So for multi-class, due to the rounding nature of feature contributions towards each class, the sum of probabilities for a particular class typically don't add up to the total value. Also, all features are not shown, so the smaller feature contributions otherwise add up as well. Take the example in github.com/marcotcr/lime and see that the contributions don't actually add up. So your observation is fine. Hope this helps. Thanks and stay tuned!
I have a problem when I tried to use foce_plot for multiple Samples. "NotImplementedError: matplotlib = True is not yet supported for force plots with multiple samples!". Can you help-me?
Thanks for the great question. Lime operates under the assumption that every complex model is linear on a local scale. So it looks at each feature for its contribution separately assuming each feature has independent contributions. SHAP however does not have any such requirement. Hope this clarifies. Stay tuned
Thanks for the question. So LIME analyzes per record at a time. If you want relative analysis, you should use SHAP visualizations that shows relative performance per record. SHAP takes more time to generate but may work well for your case. Good luck and stay tuned!
I'm working on a project in school where I will have to use XAI to get login analytics from users in terms of successful and unsuccessful login attempts using user passwords, usual login times or device location as conditions. Please how do I go about this?
@@AIwithSohini Honestly, I need to know the XAI technique best for this project. Reason for access granted or denied on the authentication system and relayed to the system administrator via a management dashboard for analytics.
Hello, so LIME interprets feature importance in terms of regression weights. The y intercept is an outcome of the regression and does not impact interpretation of features or their weightages. The weightages are still simplifications of complex models in a linear manner so they should not be considered as strict weights. Thanks and stay tuned
Great question. Not directly. If you can use neural networks to extract a linear feature layer then that layer could serve as features, but still it will be difficult to explain each pixel location. Tensorflow has better image explaining capabilities. Hope that helps. Good luck
Hello there are several towardsdata science blogs in this topic like this one towardsdatascience.com/interpreting-image-classification-model-with-lime-1e7064a2f2e5. Hope this helps. Thanks and stay tuned
Hi mam, Thanks for your inputs. Happy to see you and your tutorials again. I believe that you are from biomedical research background. Can you share some useful links related with available biomedical databases to do research using ML/DL approaches that will be helpful.
Hello. Sure I will add more links to downloadable data in the description box. If you can be specific on what medical image domain you are interested I can suggest data accordingly. Thanks
@@AIwithSohini Hi mam, I have also came across grand challenge (grand-challenge.org/). Would you suggest this for experimental purpose and whether the annotated files provided by this repo. Kindly reply
Hello, typically SHAP features are better in explaining context within respect to other samples or reference points in time. However it is memory heavy. Hope this helps. Stay tuned
@@AIwithSohini I tried Lime with logistirc regression. But when I am trying implement on Rnn Instead of using Logistic regression, it's giving error. Does lime require Logistic regression?
@@AIwithSohini pip's dependency resolver does not currently take into account all pacakges that are installed. This behavior is the source of following dependency conflicts. flask 1.1.4 requires jinja2=2.10.1, but you have jinja2, 3.2.1 which is incompatible.
Hello and thanks for the Excellent Question! So SHAP features have a capability to (in a way) visualize clusters of data streams in sequence. Also the capability of predicting positive trends over negative trends for clusters facing bias is significantly visible using SHAP. Here is a blog on using SHAP databricks.com/blog/2019/06/17/detecting-bias-with-shap.html. If you see the example here, the representation of salary based on gender is can be visible based on the variance represented in each cluster. Hope this answers your query on bias.
![[TH] RNT vs SPG - VCT Ascension Pacific - Day 1](http://i.ytimg.com/vi/RUTkf8Detj8/mqdefault.jpg)
Waiting for more lectures on data visualization, good work
Thanks Sohini, great work
Thanks so much for the encouragement. Stay tuned
Thank you so much for this video, you saved my day! You have a new subscriber :)
Most welcome. Please stay tuned
Hi, Thank you for such a nice presentation on ML Explainability.
I am learning ML, and so far I understood how scikit-learn library and other different Python libraries have made life easier in dealing with any kind of data. Now, I am learning ML Explainability, thinking to build a project with the data available in the public. Can you suggest me which dataset/s would be more suitable for me to start with as a beginner?
Hello and thanks for following my channel. Yes biased data sets are very common in e-commerce data sets. So search for e-Commerce data sets in Kaggle or the UCI Machine Learning data sets are great as well.
Sometimes with LIME we see that probabillies of each feature contribution do not sum up to probabity of the (not predicted) class (e.g 0.47). Why?
Hello and thanks for your question. So lime is a way to explain the weightages or coefficients of non linear models as a linear model would do. That’s why the feature contributions are estimates only which makes them rounded off in certain situations. Hope this helps and stay tuned
@@AIwithSohini so if the predicted probability of a certain class in a multiclass classification problem (with LIME) is 0.03 and the sum of the contributions is more than 0.10 (as in my example), the reason is because the digits are rounded?
Hello there. So for multi-class, due to the rounding nature of feature contributions towards each class, the sum of probabilities for a particular class typically don't add up to the total value. Also, all features are not shown, so the smaller feature contributions otherwise add up as well. Take the example in github.com/marcotcr/lime and see that the contributions don't actually add up. So your observation is fine. Hope this helps. Thanks and stay tuned!
I have a problem when I tried to use foce_plot for multiple Samples. "NotImplementedError: matplotlib = True is not yet supported for force plots with multiple samples!". Can you help-me?
Please search for this error in stackoverflow. This is not something I have encountered. It might just be a typo. Hope that helps.
@Ai with Sohini maam are we assuming that features are independent of each other before passing it to lime or shap??
Thanks for the great question. Lime operates under the assumption that every complex model is linear on a local scale. So it looks at each feature for its contribution separately assuming each feature has independent contributions. SHAP however does not have any such requirement. Hope this clarifies. Stay tuned
Hello Mam Instead of finding one row analysis in LIME how can we find same analysis for multiple rows at a same time is it possible?
Thanks for the question. So LIME analyzes per record at a time. If you want relative analysis, you should use SHAP visualizations that shows relative performance per record. SHAP takes more time to generate but may work well for your case. Good luck and stay tuned!
@@AIwithSohini Thank you for your guidance
I'm working on a project in school where I will have to use XAI to get login analytics from users in terms of successful and unsuccessful login attempts using user passwords, usual login times or device location as conditions. Please how do I go about this?
Hello there. Thats an interesting project. Please share details of the tools you are using and I can help out with tool selection. Good luck!
@@AIwithSohini Honestly, I need to know the XAI technique best for this project. Reason for access granted or denied on the authentication system and relayed to the system administrator via a management dashboard for analytics.
Hello, Really enjoyed your video. Can you explain what the term Intercept mean in LIME.
Hello, so LIME interprets feature importance in terms of regression weights. The y intercept is an outcome of the regression and does not impact interpretation of features or their weightages. The weightages are still simplifications of complex models in a linear manner so they should not be considered as strict weights. Thanks and stay tuned
is LIME suitable for image regression?
Great question. Not directly. If you can use neural networks to extract a linear feature layer then that layer could serve as features, but still it will be difficult to explain each pixel location. Tensorflow has better image explaining capabilities. Hope that helps. Good luck
How to export Lime output (show_in_notebook) into image?
7:59 what is "right"
hi, is there any video for LIME and SHAP on image dataset? IT will be helpful
Hello there are several towardsdata science blogs in this topic like this one towardsdatascience.com/interpreting-image-classification-model-with-lime-1e7064a2f2e5. Hope this helps. Thanks and stay tuned
Hi mam, Thanks for your inputs. Happy to see you and your tutorials again. I believe that you are from biomedical research background. Can you share some useful links related with available biomedical databases to do research using ML/DL approaches that will be helpful.
Hello. Sure I will add more links to downloadable data in the description box. If you can be specific on what medical image domain you are interested I can suggest data accordingly. Thanks
@@AIwithSohini Thanks for your kind response mam, Waiting for your links and video series. Have a wonderful day
@@venkatesanr9455 I have added data links to the comments. Hope they are helpful. Thanks and stay tuned!
@@AIwithSohini Thanks a lot, Mam and huge respect for your kind response. Also, waiting for your video series
@@AIwithSohini Hi mam, I have also came across grand challenge (grand-challenge.org/). Would you suggest this for experimental purpose and whether the annotated files provided by this repo. Kindly reply
Which one is better for RNN? Lime or Shap?
Hello, typically SHAP features are better in explaining context within respect to other samples or reference points in time. However it is memory heavy. Hope this helps. Stay tuned
@@AIwithSohini I tried Lime with logistirc regression. But when I am trying implement on Rnn Instead of using Logistic regression, it's giving error. Does lime require Logistic regression?
@@fahimhasan4905 Try this one github.com/marcotcr/lime/blob/master/doc/notebooks/Lime%20with%20Recurrent%20Neural%20Networks.ipynb
Hi Sohini
I am facing a few issues.
Hello. Sure thing. Please mention your issues here and I can try to help out.
@@AIwithSohini pip's dependency resolver does not currently take into account all pacakges that are installed. This behavior is the source of following dependency conflicts. flask 1.1.4 requires jinja2=2.10.1, but you have jinja2, 3.2.1 which is incompatible.
Video should be called "and" not "vs"
Done. Thanks so much for the input. Please do stay tuned
how to find biasness in model using shap / lime ??
Hello and thanks for the Excellent Question! So SHAP features have a capability to (in a way) visualize clusters of data streams in sequence. Also the capability of predicting positive trends over negative trends for clusters facing bias is significantly visible using SHAP. Here is a blog on using SHAP databricks.com/blog/2019/06/17/detecting-bias-with-shap.html. If you see the example here, the representation of salary based on gender is can be visible based on the variance represented in each cluster. Hope this answers your query on bias.
7:59 what is "right"