Видео

Haha will stick to RUclips for now :)

could someone help me

Hi Mustafa, it's not possible to solve your problem based on the information you provided. What package are you using? Are there any other example of where SHAP has been implemented for that package?

Great questions! SHAP is used to estimate Shapley values. So it inherits properties that are useful for understanding and comparing models. See these videos on the theory behind Shapley values for ML: ruclips.net/video/UJeu29wq7d0/видео.html ruclips.net/video/b9qqbFudVhI/видео.html I think LIME is not used in this way for two reasons: 1) it is slower than SHAP (especially TreeSHAP) to get individual feature weights. This can make it time-consuming to aggregate and analyse many instances. 2) There are no built-in functions to create aggregated plots for LIME like you see in the SHAP package. In general, LIME has fallen out of favour due to the solid theory and speed of SHAP values. So perhaps the creators have decided not to develop the method further.

Hi Eli, yes I agree that there are better approaches if you want to analyse the "potential" of different channels. You could use metrics like correlation to avoid training a bunch of models (each model would take me about 10 to 16 hours to train). However, these would not tell you which channels the trained model is using to do segmentation. "Importance" is about what is important to a model and not to the problem in general. The method you described could not be used to explain a trained model as you would be evaluating different models. Lastly, PCI can thought of as an extension of permutation feature importance (PFI). This is a commonly accepted method for explaining models built on tabular data. In the same way, correlation or models built on individual features would not provide the same information as PFI scores.

@@adataodyssey Oh i see. Thanks!

Thanks! (1) No, SHAP is not model so it is not supervised or unsupervised learning algorithm. It is a method used to explain a model. (2) I'm not familiar with this usecase. But SHAP can be used whenever you have input variables, a function and output. Where you want to explain the contributions each of the input variables to the output. In the context of predictive machine learning, you can use SHAP to explain the contributions of each model feature to a prediction.

@@adataodyssey Thank you for your prompt reply and advice. Best wishes.

Hi, good question. You may want to include additional features that capture this global information. For example the average value for one of the features in the month or year before the reading. This will allow you to understand if the current feature value is high/low relative to the average. However, in this case, IsolationForest (IF) may not be the best model. On further investigation, IF cannot model interactions. You would either need to explicitly add a feature that captures the interaction. Like the ratio of the current reading to the average. Or use a model for anomaly detection that does capture interactions like an auto encoder.

@@maxpain6666 no problem!

For classification problems, shapley values are interpreted in terms of log odds. See this video for more details: ruclips.net/video/2xlgOu22YgE/видео.html

No problem!

You can read my conference paper on the topic: arxiv.org/abs/2405.11500 Permutation with superpixels is more aimed at explaining important regions/ groups of pixels. Like how they are used in SHAP or occlusion.

Wow, my first super thanks. Much appreciated Mary

Thanks Danish :)

That's great! I guess this is the end of CatBoost then 😅

Thanks Joseph :)

Here you go: github.com/conorosully/SHAP-tutorial/blob/main/src/shap_tutorial.ipynb

Hi Sukhsehaj, good question. I discuss the application to ML in this video: ruclips.net/video/b9qqbFudVhI/видео.html To summarise, you train one model with all features. To get the contribution of different feature sets, you "marginalise" over the features that are not in the set.

Sure! You can find it here: github.com/conorosully/SHAP-tutorial/blob/main/src/additional_resources/IsolationForest.ipynb

No problem Vic! Will be out with some new content soon :)

No problem Pablo!

No problem! They are called Accumulated Local Effects (ALE) Plots. I have a video on the topic: ruclips.net/video/5KCA1FMy6U4/видео.html

Thanks 😁 ​@@adataodyssey

Hi Togy, thanks! Yes, that's no problem at all. My linkedin page is: www.linkedin.com/in/conor-o-sullivan-14b267142/

These are the coalition values (i.e. prize money) if only the 2 players were in a team. To get these, I said "we need to go back in time a few more times". This is so that each alternative team could play the game and win prize money. Obviously, this is not possible. In reality, we must estimate these values. So, the actual prize money of $10000 must be split across the 3 players. However, to find a fair way to divide the money we need to know how much players would have won in alternative teams of 2 or less members.

@@adataodyssey thank you very much for the detailed answer. the marginal coalition values are calculated in advance right ?

@@yaelefrat7677 They will need to be determined in advanced. For some applications they can be observed or estimated but not calculated. For other applications, like machine learning they are calculated.

Thanks Yash!

My pleasure :)

You should receive a coupn code in your mail. Let me know if you don't get it!

Thanks Jhaoui!

No problem Zahra! I'm glad you found it useful

No problem :)

No problem Panh!

Thanks Tim! It's good to hear I am covering relevant topics :)

Thank you! The hope is that someone out there will find it useful :)

@@adataodyssey you're a superhero without a cape.

@@adataodyssey this is my the topic of my Master's so you're really helping lol

@@Daniel-RedTsar I'm glad I coud help. Good luck with the masters!

Embrase the struggle! Will you will learn more through trying to solve it :)

Often, even if you have highly correlated features, SHAP will still work. It is just important to keep in mind that it may have problems if you do have highly correlatated features. In this case, you just need to confirm the results from shap using a method that is robust to them like ALEs or simple data exploration methods.

@@civilengineeringonlinecour7143 thanks glad it could help!

@@civilengineeringonlinecour7143 I’m glad you found it useful!

@@adeauy2294 The plots should be the same if you train a NN on tabular data. However, I’ve had a lot of trouble trying to get the package to work with PyTorch. I’m not sure about Keras but I expect you are having similar problems.

@@brenoingwersen784 The problem is if you have a categorical feature with many categories (say 10), you will have 10 dummy features after encoding. This means you will have 10 SHAP values for the categorical feature making it difficult to understand the overall effect of that feature. You can solve this by adding the SHAP values for each dummy feature or using catboost.

@@cauchyschwarz3295 This is a bit confusing! But no, you will only have one model (the one you want to explain). You marginalise the prediction function (I.e. model) over the features that are not in set S. You do not have to retrain a new model by excluding the features that are not in S.

I think there's a subtlety here in the application. In the theoretical model the goal is to calculate a fair payout for an individual feature based on how much it contributes to building a good model. Good models are measured by how well they predict things. So, we would want to think about how much the feature contributes to a reduction in model error, and you would want to train a model with and without a given feature to figure this out. (But instead we use other methods for determining which features are good that are easier to compute.) In the use case in this video (and what is the norm in ML), they are using Shapley to explain how a feature contributed to the model prediction, irregardless of how good the model actually is. This is helpful because Shapley still has desireable properties like additivity in this use case. If you trained a new model without this feature, it wouldn't answer the question of how much the feature contributed to the prediction value in the old model.