To learn more about one common way to create histograms of continuous variables, see: journals.plos.org/plosone/article?id=10.1371/journal.pone.0087357 To learn more about.... R-squared = ua-cam.com/video/2AQKmw14mHM/v-deo.html Entropy = ua-cam.com/video/YtebGVx-Fxw/v-deo.html To learn more about Lightning: lightning.ai/ Support StatQuest by buying my books The StatQuest Illustrated Guide to Machine Learning, The StatQuest Illustrated Guide to Neural Networks and AI, or a Study Guide or Merch!!! statquest.org/statquest-store/
Super! I have been struggled between copula, mutual information, etc. for a while, that is exactly what I am looking for! Thank you, Josh! This video is really helpful!
Dude, you don't even know me, and I don't really know you either, but oh boyy, I fucking love you. Thank you. One day I will teach people just like you do.
It would be awesome if there were links to "if you are not familiar with XYZ, check out the quest", for noobs trying figure out what we don't know. Keep up the great work!
Those links are at the bottom of the description, but I'll also add them here: R-squared = ua-cam.com/video/2AQKmw14mHM/v-deo.html Entropy = ua-cam.com/video/YtebGVx-Fxw/v-deo.html
Seriously though, I think the KL divergence is worth a mention here. Mutual information appears to be the KL divergence between the actual (empirically derived) joint probability mass function, and the (empirically derived) probability mass function assuming independence. I know that's a lot of words, but my brain can't help seeing these relationships.
Hello, that's a great video and it has helped me understand a lot about Mutual Information as well as your other video about entropy. I do have a question. At 11:13 the answer you get after calculation is 0.5004 and it is explained that it is close to 0.5. However when I do the math (( 4 ÷ 5 ) × log ( 5 ÷ 4 ) + ( 1 ÷ 5 ) × log( 5 ) ) the answer I get is 0.217322... Am I missing something? Because from what I understood, the closer you get to 0.5, the better it is but it is not confirmed by my other examples. Is there a maximum to mutual information? Thank you for your video.
The problem is that you are using log base 10 instead of the natural log (log base 'e'). I talk about this at 8:07 and in this other video: ua-cam.com/video/iujLN48gumk/v-deo.html
1) based on what to choose the number of bins? Does larger number of bins gives lesser mutual information? 2) what if the label (output value) is numerical? Thank in advance
1) Here's how a lot of people find the best number (and width) of the bins: journals.plos.org/plosone/article?id=10.1371/journal.pone.0087357 2) Then you make a histogram of the label data.
so here does it means that we are comparing two variables, one is feature and one is output and the output is taken from test data? and basically we are tuning the model and we are using mutual information just to know which of the features are more useful to tune our model to get more accurate predictions? and after this we check our tuned model for the test set? and why do we want to reduce the attributes? do we do it because the less attributes will do the fast calculations and we can train our data in less time?
That's the main idea. There are a lot of reasons you might want to reduce the number of variables in your model. 1) sometimes collecting data can be very expensive 2) fewer variables can mean we need less data to fit the model.
3 more things: 1- it would have been great if you could make a comparison with correlation too here, 2- discuss the minimum and maximum value of the MI, 3- the intuition of this specific formula
Thanks! I'm not really sure you can compare Mutual Information to correlation because correlation doesn't work at all with discrete data. I mention this at 1:20.
Presumably - pretty much everything is affected by imbalanced data. This is because you have a much better estimate one class and a much worse estimate for the other.
hi, what will be the base of the logarithm when calculating entropy. I believe it was mentioned in the entropy video that for 2 outputs(yes/no or heads/tails) the base of the logarithm will be two. Is there any generalization to this statement?
is there a good and stable way to calculate mutual information for numeric variables *where the binning is not good*, e.g. highly skewed distributions where the middle bins are very different from the edge bins?
Off topic question...but will chatgpt replace us as data scientists/analysts/ statisticians. I just discovered it tonight and it blew me away. I basically learned html and css in a day with it. Im worried it will massively reduce jobs in our field. I did a project that would normally take all day in a few minutes...scary stuff.
Well, if you really want his opinion, watch the AI Buzz #1 Josh uploaded three weeks ago. It’s in this channel. As for my opinion, obviously nobody knows yet, but it will soon be a new ground-level for anybody else. For some that all they can do is basic things ChatGPT does far better, they are in danger; for others that can make more values out of ChatGPT (or any tools to come), they are in far better shape. Which do you think you and fellow data scientists are? And even for the basic stuffs, there should be at least someone to check whether the ChatGPT has done some absurd work or not, right? Maybe at least for a few years or so.
@@statquest thank you! This is great. Im also starting my first job today post college as a research data specialist! Your videos always helped me throughout my data science bachelors, so thank you!
Is it weird that my prof. gave me the mutual information equation as one that uses entropy? We were given "I(A; B) = H(B) - sum_b P(B = b) * H(A | B = b)" with no mention of the equation you showed in this video
That is odd. Mutual information can be derived from the entropy of two variables. It is the average of how the surprise in one variable is related to the surprise in another. However, this is the standard formula. See: en.wikipedia.org/wiki/Mutual_information
Hey, so what if our dependent variable ( here, loves troll 2) is continuous? Can we use Mutual information in that case? by binning aren't we just converting it into a categorical variable?
Josh, thank you for the awesome easily digestible video. One question. Is there any specific guideline about binning the continuous variable? I'm fairly certain that depending on how you split it (how many bins you choose and how spread they are) the result might be different.
To learn more about one common way to create histograms of continuous variables, see: journals.plos.org/plosone/article?id=10.1371/journal.pone.0087357
@@statquest Josh, thank you for the link, but I guess I formulated my question incorrectly. The question was about not creating the histogram but actually choosing the bins. You split your set in 3 bins. Why 3? Why not 4 or 5? Would the result change drastically if you split in 5 bins? What if the distribution of the variable you are splitting is not normal or uniform? Etc
@@wowZhenek When building a histogram, choosing the bins is the hard part, and that is what that article describes - a special way to choose the number and width of bins specifically for Mutual Information. So take a look. Also, because we are using a histogram approach, it doesn't matter what the underlying distribution is. The histogram doesn't make any assumptions.
@@statquest oh, yeah, I didn't look inside the URL you gave because your described it as "one common way to create histograms of continuous variables" which seemed very much distant from what I was actually asking about. Now that I checked the link, damn, what a comprehensive abstract. Thank you very much!
this is a nice tutorial and with different useful scenarios. But I didn't completely grasp the intuition of something never changing telling nothing about something that does. I understand it mathematically but hoping for a more intuitive explanation, because even if something does not change, there are some matches between the features.
Say like I ask a bunch people what is their favorite color is and how old they are. Some of the people are young, some are middle aged and some are old, but everyone loves the color green. Now, if I told you that someone in that group loved the color green, what would that tell you about that person's age? Nothing. Since everyone loves green (it never changes) it can't differentiate between young, middle aged and old people.
@@jozefinagramatikova4889 When used with linear regression, then yes. However, R-squared can be applied to any model, even models that make non-linear fits, and in that case, it can evaluate a non-linear relationship.
@@statquest Thank you very much! So, is Mutual Information used more often compared to R^2 for feature selection (when we don't have categorical features) and why?
@@statquest Is MI then somehow influenced by the size of the data or the number of categories? The video seems to suggest it should be around 0.5 for perfectly shared information (at least in this example). With discrete data using 15 bins I get some values close to 1. Thanks for these great videos.
@@Chuckmeister3 Interpretation from coding theory (natural log replaced by log to base 2): Mutual information I(X;Y) is the amount of bits wasted if X and Y are encoded separately instead of jointly encoded as vector (X,Y). Statement holds on average and only asymptotically, i.e. for optimal entropy coding (e.g. arithmetic encoder) with large alphabets (asymptotically for size -> oo). It's the amount of information shared by X and Y measured in bits. Mutual information can become arbitrarily large, depending on the size of the alphabets of X and Y (and the distribution p(x,y) of course). But it can't be greater than the separate entropies H(X) and H(Y), respectively the minimum of both. You can think of I(X;Y) as the intersection of H(X) and H(Y). ps: I think the case of perfectly shared information is if there's a (bijective) function connecting each symbol of X with each symbol of Y, so that the relation between X and Y becomes deterministic. In that case H(X)=H(Y)=I(X;Y). The other extreme is X and Y being statistically independent: In that case I(X;Y) = 0.
If I want to calculate the correlation between Likes Popcorn and Likes Troll 2, can I use something like Chi2? Similarly between Height bins and Likes Troll 2. What's the advantage of calculating the Mutual Information?
The advantage is that we have a single metric that works on both continuous, discrete and mixed variables and we don't have to make any assumptions about the underlying distributions.
I was always interested how should we think if we want to invent such a technique. Imean ok, lets say I "suspect" that the probabilities here should do the job, and say my goal is to get at the end of a day some "flag" from 0 to 1 which indicates the strenght of a relationship, but how should I think on, to deside like what comes to denominator vs nominator, when use log etc. There should be something like an "thinking algorithm" P.s Understanding this will be very helpfull in understanding the existing fancy formulas
To learn more about one common way to create histograms of continuous variables, see: journals.plos.org/plosone/article?id=10.1371/journal.pone.0087357
To learn more about....
R-squared = ua-cam.com/video/2AQKmw14mHM/v-deo.html
Entropy = ua-cam.com/video/YtebGVx-Fxw/v-deo.html
To learn more about Lightning: lightning.ai/
Support StatQuest by buying my books The StatQuest Illustrated Guide to Machine Learning, The StatQuest Illustrated Guide to Neural Networks and AI, or a Study Guide or Merch!!! statquest.org/statquest-store/
Thank u daddy stat quest for carrying me through my university course
Ha! :)
I am binge-watching this series. Very clear and concise explanations for every topics given in the most interesting way!
Glad you like them!
Same here!
Superb!!! I recommend this channel to everyone.
Thanks!
Mesmerizing! U are a beacon of hope for us struggling engineers here in China xxx
Thanks!
Thank you for being a content creator
Thanks!
Not just a creator of any content either. A creator of *exceptional* content!
Mutual information, clearly explained? More like "Magnificent demonstration, you deserve more fame!" 👍
Thanks! 😃
Super! I have been struggled between copula, mutual information, etc. for a while, that is exactly what I am looking for! Thank you, Josh! This video is really helpful!
Glad it was helpful!
Superb explanation! Your channel is great!
Glad you think so!
Dude, you don't even know me, and I don't really know you either, but oh boyy, I fucking love you. Thank you. One day I will teach people just like you do.
Thanks! :)
Entropy === The expectation of the surprise!!! I'll never look at this concept the same again
bam! :)
Awesome stuff, Josh. Thank you!
My pleasure!
OMG i never see this channel, how many hours would be saveeddd.. new subs here, thanks alottt for ur vids
Welcome!
Your explanations are awesome!
Glad you like them!
I love this video. Simple and clear.
Thanks!
It would be awesome if there were links to "if you are not familiar with XYZ, check out the quest", for noobs trying figure out what we don't know. Keep up the great work!
Those links are at the bottom of the description, but I'll also add them here:
R-squared = ua-cam.com/video/2AQKmw14mHM/v-deo.html
Entropy = ua-cam.com/video/YtebGVx-Fxw/v-deo.html
❤ There's a lot of junk appended by yt to the description, I had to look hard to find it just now
Great explanation, thank you! ❤🔥
Glad it was helpful!
Just started Learning ML, am assured now that the journey would be smooth with this channel
Good luck! :)
Great stuff. As always.
Thank you very much! :)
An interesting explanation and nice sence of humor 👍
Thank you!
you are the best god sent really stay blessed
Thank you!
Great job! Love it!
Liking my own comment to double like your video :)
Double bam! :)
that was quite useful brother thanks
Thanks!
Someone hit me on the head with a club, and now I'm good at stats. That's what they call... bam bam.
ha! :)
Seriously though, I think the KL divergence is worth a mention here.
Mutual information appears to be the KL divergence between the actual (empirically derived) joint probability mass function, and the (empirically derived) probability mass function assuming independence.
I know that's a lot of words, but my brain can't help seeing these relationships.
One day I hope to do a video on the KL divergence.
The calculation at 8:27 seems incorrect. I plugged it into a calculator and got 0.32. The log is base 2 right?
At 8:07 I say that we are using log base 'e'.
Thank you so mcuh! It is really helpful. I really hope you can explain KL divergence in the next video.
I'll keep that in mind.
Amazing as always!!!
Thank you!
Love it, thanks!
Thank you!
Two sigmas are like two for loops, such that, for every index of outer Sigma, the inner sigmaales a complete iteration.
bam!
Thank youuuu. you explain everything clearly
Glad it was helpful!
I love this channel
BAM! :)
@@statquest lol, very on-brand too.
You got a like just for the musical numbers!
bam!
Hello, that's a great video and it has helped me understand a lot about Mutual Information as well as your other video about entropy. I do have a question.
At 11:13 the answer you get after calculation is 0.5004 and it is explained that it is close to 0.5. However when I do the math (( 4 ÷ 5 ) × log ( 5 ÷ 4 ) + ( 1 ÷ 5 ) × log( 5 ) ) the answer I get is 0.217322... Am I missing something? Because from what I understood, the closer you get to 0.5, the better it is but it is not confirmed by my other examples. Is there a maximum to mutual information?
Thank you for your video.
The problem is that you are using log base 10 instead of the natural log (log base 'e'). I talk about this at 8:07 and in this other video: ua-cam.com/video/iujLN48gumk/v-deo.html
@@statquest Thank you for your answer. That explains a lot.
I have same doubt, when both columns are equal it says mutual info is 0.5 then what is maximum value of mutual info and in which scenario ?
1) based on what to choose the number of bins? Does larger number of bins gives lesser mutual information?
2) what if the label (output value) is numerical?
Thank in advance
1) Here's how a lot of people find the best number (and width) of the bins: journals.plos.org/plosone/article?id=10.1371/journal.pone.0087357
2) Then you make a histogram of the label data.
so here does it means that we are comparing two variables, one is feature and one is output and the output is taken from test data? and basically we are tuning the model and we are using mutual information just to know which of the features are more useful to tune our model to get more accurate predictions? and after this we check our tuned model for the test set? and why do we want to reduce the attributes? do we do it because the less attributes will do the fast calculations and we can train our data in less time?
That's the main idea. There are a lot of reasons you might want to reduce the number of variables in your model. 1) sometimes collecting data can be very expensive 2) fewer variables can mean we need less data to fit the model.
Your explanations are alway awesome! I wonder how to explain Normalized Mutual Information?
I believe it's just a normalized version of mutual information (so scale it to be a value between 0 and 1).
This is great! Do you know if you can interpret a NMI value in percentages, something like 7% of information overlaps, or 7% of group members overlap?
Keep it up. Great content
Thank you!
3 more things: 1- it would have been great if you could make a comparison with correlation too here, 2- discuss the minimum and maximum value of the MI, 3- the intuition of this specific formula
Thanks! I'm not really sure you can compare Mutual Information to correlation because correlation doesn't work at all with discrete data. I mention this at 1:20.
Amazing as always! Any update on the transformer video?
Still working on it.
thankss joshh 😍😍 in 1:30 since the response variable is not continuous and takes on 0 or 1(yes/no) can we model it with logistic regression?
Yep!
In case of continuous variables how to decide the number of bins and the boundaries?
It probably depends on the dataset. Usually with things like that I like to plot histograms to make decisions.
Fire🔥🔥🔥
BAM! :)
Hi, thank you Josh. I have one question. Does MI score is affected by imbalanced data?
Presumably - pretty much everything is affected by imbalanced data. This is because you have a much better estimate one class and a much worse estimate for the other.
hi, what will be the base of the logarithm when calculating entropy. I believe it was mentioned in the entropy video that for 2 outputs(yes/no or heads/tails) the base of the logarithm will be two. Is there any generalization to this statement?
Unless there is a specific reason to use a specific base for the log function, we use log base 'e'.
awesome!!! Now waiting for a video on Chi2 Test of Independence.
I'll keep that in mind.
is there a good and stable way to calculate mutual information for numeric variables *where the binning is not good*, e.g. highly skewed distributions where the middle bins are very different from the edge bins?
Hmm... off the top of my head, I don't know, but I wouldn't be surprised if there was someone out there publishing research papers on this topic.
you are a genius
:)
Can we have videos about all the gazillion hypothesis tests available!!
I'll keep that in mind.
Off topic question...but will chatgpt replace us as data scientists/analysts/ statisticians. I just discovered it tonight and it blew me away. I basically learned html and css in a day with it. Im worried it will massively reduce jobs in our field. I did a project that would normally take all day in a few minutes...scary stuff.
Well, if you really want his opinion, watch the AI Buzz #1 Josh uploaded three weeks ago. It’s in this channel.
As for my opinion, obviously nobody knows yet, but it will soon be a new ground-level for anybody else. For some that all they can do is basic things ChatGPT does far better, they are in danger; for others that can make more values out of ChatGPT (or any tools to come), they are in far better shape. Which do you think you and fellow data scientists are?
And even for the basic stuffs, there should be at least someone to check whether the ChatGPT has done some absurd work or not, right? Maybe at least for a few years or so.
Just out of curiosity how did you learn HTML and CSS in a day ?
And what's specific task that you solved
I didnt think ChatGPT is that impressive afterall. Makes so many mistakes is not able to do really complicated stuff. Totally overhyped!
See: ua-cam.com/video/k3b9Mvtt6lU/v-deo.html
@@statquest thank you! This is great. Im also starting my first job today post college as a research data specialist! Your videos always helped me throughout my data science bachelors, so thank you!
you are amazing
Thank you!
Excellent content as always!
Much appreciated!
Can't we use correlation factor instead of Mutual information for continuous variable?
If you have continuous data, use R^squared.
Can this mutual information value be greater than 0.5, I mean closer to 1??
In theory the range of possible values goes from 0 to positive infinity.
Is it weird that my prof. gave me the mutual information equation as one that uses entropy? We were given "I(A; B) = H(B) - sum_b P(B = b) * H(A | B = b)" with no mention of the equation you showed in this video
That is odd. Mutual information can be derived from the entropy of two variables. It is the average of how the surprise in one variable is related to the surprise in another. However, this is the standard formula. See: en.wikipedia.org/wiki/Mutual_information
Hey, so what if our dependent variable ( here, loves troll 2) is continuous? Can we use Mutual information in that case? by binning aren't we just converting it into a categorical variable?
You could definitely try that.
Can you please make a video on Latent Dirichlet Allocation
I'll keep that in mind! :)
Josh, thank you for the awesome easily digestible video. One question. Is there any specific guideline about binning the continuous variable? I'm fairly certain that depending on how you split it (how many bins you choose and how spread they are) the result might be different.
To learn more about one common way to create histograms of continuous variables, see: journals.plos.org/plosone/article?id=10.1371/journal.pone.0087357
@@statquest Josh, thank you for the link, but I guess I formulated my question incorrectly. The question was about not creating the histogram but actually choosing the bins. You split your set in 3 bins. Why 3? Why not 4 or 5? Would the result change drastically if you split in 5 bins? What if the distribution of the variable you are splitting is not normal or uniform? Etc
@@wowZhenek When building a histogram, choosing the bins is the hard part, and that is what that article describes - a special way to choose the number and width of bins specifically for Mutual Information. So take a look. Also, because we are using a histogram approach, it doesn't matter what the underlying distribution is. The histogram doesn't make any assumptions.
@@statquest oh, yeah, I didn't look inside the URL you gave because your described it as "one common way to create histograms of continuous variables" which seemed very much distant from what I was actually asking about. Now that I checked the link, damn, what a comprehensive abstract. Thank you very much!
It seems information gain (defined via entropy) and mutual information are the same thing?
They are related, but not the same thing. For details, see: en.wikipedia.org/wiki/Information_gain_(decision_tree)
@@statquest Thanks, I'll check it out. And also thanks for all the videos. It's an incredible resource you've produced.
this is a nice tutorial and with different useful scenarios.
But I didn't completely grasp the intuition of something never changing telling nothing about something that does. I understand it mathematically but hoping for a more intuitive explanation, because even if something does not change, there are some matches between the features.
Say like I ask a bunch people what is their favorite color is and how old they are. Some of the people are young, some are middle aged and some are old, but everyone loves the color green. Now, if I told you that someone in that group loved the color green, what would that tell you about that person's age? Nothing. Since everyone loves green (it never changes) it can't differentiate between young, middle aged and old people.
6:18 not small bam, big bam... thank you very much...
BAM!!! :)
So when we don't have categorical features can we just use R^2?
Yep
@@statquest But doesn't R^2 show only linear relationship?
@@jozefinagramatikova4889 When used with linear regression, then yes. However, R-squared can be applied to any model, even models that make non-linear fits, and in that case, it can evaluate a non-linear relationship.
@@statquest Thank you very much! So, is Mutual Information used more often compared to R^2 for feature selection (when we don't have categorical features) and why?
@@jozefinagramatikova4889 If you don't have categorical features, I think R^2 is more popular.
Que Top. Dublado em português
Muito obrigado! :)
maybe next video on this: KL divergence
It's on the list.
DOUBLE BAM!!
Thanks!
What does it mean if mutual information is above 0.5? If 0.5 is perfectly shared information...
As you can see in the video, perfectly shared information can have MI > 0.5. So 0.5 is not the maximum value.
@@statquest Is MI then somehow influenced by the size of the data or the number of categories? The video seems to suggest it should be around 0.5 for perfectly shared information (at least in this example). With discrete data using 15 bins I get some values close to 1.
Thanks for these great videos.
@@Chuckmeister3 Yes, the size of the dataset matters.
@@Chuckmeister3 Interpretation from coding theory (natural log replaced by log to base 2): Mutual information I(X;Y) is the amount of bits wasted if X and Y are encoded separately instead of jointly encoded as vector (X,Y). Statement holds on average and only asymptotically, i.e. for optimal entropy coding (e.g. arithmetic encoder) with large alphabets (asymptotically for size -> oo). It's the amount of information shared by X and Y measured in bits. Mutual information can become arbitrarily large, depending on the size of the alphabets of X and Y (and the distribution p(x,y) of course). But it can't be greater than the separate entropies H(X) and H(Y), respectively the minimum of both. You can think of I(X;Y) as the intersection of H(X) and H(Y).
ps: I think the case of perfectly shared information is if there's a (bijective) function connecting each symbol of X with each symbol of Y, so that the relation between X and Y becomes deterministic. In that case H(X)=H(Y)=I(X;Y). The other extreme is X and Y being statistically independent: In that case I(X;Y) = 0.
awesome
Thanks!
this is cool
Thanks!
i'm more of a 'Goblin 3: the frolicking' man myself
bam!
It's like FoodWishes for stats
:)
If I want to calculate the correlation between Likes Popcorn and Likes Troll 2, can I use something like Chi2? Similarly between Height bins and Likes Troll 2. What's the advantage of calculating the Mutual Information?
The advantage is that we have a single metric that works on both continuous, discrete and mixed variables and we don't have to make any assumptions about the underlying distributions.
tiny bam
:)
Ummm I know I have a cold right now but did anyone only hear an Italian girl speaking ?
?
puop poopup pooh
:)
small bam = "bamsito"
Ha! :)
2:11 baam
8:48 double baaaam
9:28 tiny baaaam
ha! :)
If I could I'd kiss you on the mouth, wish you did a whole playlist about data compression
Ha! I'll keep that topic (data compression) in mind.
Great content. But just don't sing, you're not up to that.
Noted! :)
i will fite you if you tell daddy stat quest what to do what not to do
I was always interested how should we think if we want to invent such a technique. Imean ok, lets say I "suspect" that the probabilities here should do the job, and say my goal is to get at the end of a day some "flag" from 0 to 1 which indicates the strenght of a relationship, but how should I think on, to deside like what comes to denominator vs nominator, when use log etc. There should be something like an "thinking algorithm"
P.s
Understanding this will be very helpfull in understanding the existing fancy formulas
I talk more about the reason for the equation in my video on Entropy: ua-cam.com/video/YtebGVx-Fxw/v-deo.html
that small bam
:)