This is another fun video! I took my time trying to edit this one so it's easy to follow. There are technical details throughout the video, but hoping this one is well paced with adequate pauses to digest material. Also removed the background music for this one since it might be distracting. Please let mw know how this video pans out for you in the comments. Would appreciate some feedback as I'm trying make challenging content easy to understand. Thanks so much for the support!
To be Precise, I would express divergence in terms of conditional probabilities (Prob[Cured|Treated]-Prob[Cured|NonTreated])^2+(Prob[NotCured|Treated]-Prob[Not Cured|NonTreated])^2. In the video the definition of divergence is not same as stated above. It was defined in terms of joint probability instead of conditional probability. Please correct me , if I am wrong...
Very clear video and well-explained. However, you didn't discuss on very important issues in the causal trees, such as overlap assumption, when we cannot have an inference of the counter factual due to low overlap in some regions of the feature space
Very clear and informative video, thank your for sharing it but i have a small question about divergence formula . what if we have 10 persons , 5 treated and not cured also the another 5 not treated but cured , do u notice something ? it will have a high divergence despite that it is a horrible situations which is mean we need to ignore this split , but how whenever the divergence is too high in this case ?
Great video! I have a question. While investigating the individual level (cognitive and non-cognitive traits) and the family level factors (parental education, income, marital status, conflict with the child, warmth) that affect adolescent academic performance measured in terms of GPA, I wanted to use causal trees. Now, all my covariates and the outcome are continuous and there is no "treatment" that I am providing. In such a case, how do I use and interpret this method? I just have a set of predictors based on previous literature. Thanks in advance!
I have a video titled "Boosting - EXPLAINED" I made maybe 2 years ago. I explain many forms of boosting and it's evolution. I think the last few minutes should be dictated to Xgboost.
Very clear and informative video, thank you for making it ! After watching it, this question came up with me: in the case of “ordinary, non-causal decision trees the random forest was developed. A single decision tree is too sensitive to small changes in the data upon which it is trained; training a set of decision trees on random subsets of the data and on random subsets of the input variables in the case of random forests proved to be a good to be a good way of circumventing that brittleness of the single decision tree ! But is this also the case for causal decision trees? Does there exist Causal Forests to fight the same problem?
Excellent question! Yes. In practice, we would actually implement this as a random forest instead of using this as a decision tree because of that sensitivity issue you mentioned. In fact, i think all the resources i mentioned in the description of this video imply randomly forests are used. But to illustrate the concept if causality, i thought decision trees are easier to dissect and discuss. Overall, good thought!
@@CodeEmporium I agree that it is best to start off with causal decision trees🌲 🌲 . Causal Decision Forests are a relatively simple extension of Causal Decision Trees indeed. Keep up the good work making these highly informative videos, I appreciate them very much !
This is another fun video! I took my time trying to edit this one so it's easy to follow. There are technical details throughout the video, but hoping this one is well paced with adequate pauses to digest material. Also removed the background music for this one since it might be distracting. Please let mw know how this video pans out for you in the comments. Would appreciate some feedback as I'm trying make challenging content easy to understand. Thanks so much for the support!
I will watch that video with great interest and will react if need be !
I can't believe how very good this is ... very clearly explained, and you explained everything. Thank you!
Thank you for the compliments. Really appreciate it :)
Great video, and thanks for including the references in the notes too!
Of course! Thanks for watching!
I am a big fan of this channel... Waiting for a video on GANs from you!
Thank you!! Also, i should have a playlist of a few videos on GANs.
To be Precise, I would express divergence in terms of conditional probabilities (Prob[Cured|Treated]-Prob[Cured|NonTreated])^2+(Prob[NotCured|Treated]-Prob[Not Cured|NonTreated])^2. In the video the definition of divergence is not same as stated above. It was defined in terms of joint probability instead of conditional probability. Please correct me , if I am wrong...
Either you are right, or the probabilities in the video are wrong
Very clear video and well-explained.
However, you didn't discuss on very important issues in the causal trees, such as overlap assumption, when we cannot have an inference of the counter factual due to low overlap in some regions of the feature space
Very clear and informative video, thank your for sharing it but i have a small question about divergence formula .
what if we have 10 persons , 5 treated and not cured also the another 5 not treated but cured , do u notice something ?
it will have a high divergence despite that it is a horrible situations which is mean we need to ignore this split , but how whenever the divergence is too high in this case ?
Great vid, thank you for your work!
Thanks a ton for watching
thanks for your great explanation. I would like to ask how do we obtain the ITE after the split?
Did you do all those visuals in manim?
Nope. Straight up Camtasia :)
📈🔥
Great video! I have a question. While investigating the individual level (cognitive and non-cognitive traits) and the family level factors (parental education, income, marital status, conflict with the child, warmth) that affect adolescent academic performance measured in terms of GPA, I wanted to use causal trees. Now, all my covariates and the outcome are continuous and there is no "treatment" that I am providing. In such a case, how do I use and interpret this method? I just have a set of predictors based on previous literature. Thanks in advance!
Can you do one on XGBoost please?
I have a video titled "Boosting - EXPLAINED" I made maybe 2 years ago. I explain many forms of boosting and it's evolution. I think the last few minutes should be dictated to Xgboost.
Very clear and informative video, thank you for making it ! After watching it, this question came up with me: in the case of “ordinary, non-causal decision trees the random forest was developed. A single decision tree is too sensitive to small changes in the data upon which it is trained; training a set of decision trees on random subsets of the data and on random subsets of the input variables in the case of random forests proved to be a good to be a good way of circumventing that brittleness of the single decision tree !
But is this also the case for causal decision trees? Does there exist Causal Forests to fight the same problem?
Excellent question! Yes. In practice, we would actually implement this as a random forest instead of using this as a decision tree because of that sensitivity issue you mentioned. In fact, i think all the resources i mentioned in the description of this video imply randomly forests are used. But to illustrate the concept if causality, i thought decision trees are easier to dissect and discuss.
Overall, good thought!
@@CodeEmporium I agree that it is best to start off with causal decision trees🌲 🌲 . Causal Decision Forests are a relatively simple extension of Causal Decision Trees indeed.
Keep up the good work making these highly informative videos, I appreciate them very much !
why would I use Causal Decision tree against uplift model?
cool!
Ur god